The Textbook Has Not Aged Well
I pulled up the top 20 search results for “intelligent agent in artificial intelligence examples” last week. Seventeen of them opened with a thermostat. Three opened with a vacuum cleaner. Every single one cited Russell and Norvig's 1995 textbook definition, which is correct but has been fossilized into a kind of cargo-cult pedagogy where the same five examples cycle through blog post after blog post with no connection to anything a working engineer would recognize.
The definition itself holds up remarkably well. Wooldridge and Jennings nailed it in 1995 when they formalized the four properties that distinguish an intelligent agent from a mere program: autonomy, reactivity, proactiveness, and social ability. What has not held up is the example set. We are not in 1995 anymore. We are not even in 2020. The systems that satisfy these four properties in 2026 look nothing like what any textbook describes.
So here is what I set out to do: take the formal definition seriously, apply it rigorously, and find 12 systems that actually satisfy the Wooldridge-Jennings criteria in production. Not hypothetical agents. Not research prototypes. Systems running right now, processing real data, making real decisions, and failing in real ways when the infrastructure underneath them breaks. The thesis, if you want it upfront: intelligence without infrastructure is just a hallucinating API call.
Quick Primer: The Four Properties That Define Intelligence
Before the examples, a shared vocabulary. Every example in this article gets evaluated against these four properties. If a system does not exhibit all four, it is an agent but it is not an intelligent agent in the strict Wooldridge-Jennings sense. Most systems in production satisfy two or three. The ones that satisfy all four are the ones worth studying closely.
Autonomy
Operates without direct human intervention. Controls its own actions and internal state. A system that needs a human to press “run” every morning is not autonomous.
Reactivity
Perceives its environment and responds to changes in a timely fashion. The latency requirement varies by domain — nanoseconds for trading, seconds for customer support, minutes for research.
Proactiveness
Takes initiative toward goals. Does not merely react to stimuli but actively pursues objectives, generates sub-goals, and plans sequences of actions. The difference between a notification system and an agent.
Social Ability
Interacts with other agents — human or artificial — through communication protocols. Coordinates, negotiates, delegates, escalates.
Category 1 — Reactive Intelligence
Agents That Think in Milliseconds
Reactive intelligent agents perceive environmental changes and respond faster than any deliberative process would allow. They maintain internal models — that is what distinguishes them from simple reflex agents — but their primary mode of operation is stimulus-response at speed. The infrastructure challenge here is latency and uptime. A reactive agent that goes down for 30 seconds might as well not exist.
1. High-Frequency Trading Bots
ReactiveCitadel Securities processes roughly 27% of all U.S. equity volume. Jane Street handles over $17 billion in daily trading across 45 countries. These are not batch programs running on a schedule. They are intelligent agents perceiving market microstructure at nanosecond resolution, maintaining internal models of order book dynamics, and executing trades autonomously against competing agents. The social dimension is literal — every trade is an interaction with another agent on the other side.
Wooldridge-Jennings mapping: Autonomy — operates 24/7 without human trigger. Reactivity — responds to market data in sub-microsecond timeframes. Proactiveness — actively seeks arbitrage opportunities and adjusts position sizing. Social ability — interacts with market counterparties, exchanges, and regulatory reporting systems.
Infrastructure: Co-located servers within exchange data centers, FPGA-based execution pipelines, redundant network paths with sub-microsecond failover. A 1ms latency spike can cost millions. What breaks: Network jitter kills reactivity. Power failure kills autonomy. Stale market data models produce toxic order flow. The 2012 Knight Capital incident — $440 million lost in 45 minutes — remains the canonical infrastructure failure for reactive intelligent agents.
2. Game AI — AlphaGo and Descendants
ReactiveWhen AlphaGo defeated Lee Sedol in 2016, it evaluated roughly 200 million board positions per move using Monte Carlo tree search combined with deep neural networks. Its descendant AlphaZero generalized this approach to chess and shogi, learning superhuman play entirely through self-play with no human game data. By 2026, the same architecture powers game AI in commercial titles — Ubisoft's Ghostwriter generates NPC dialogue, and Nvidia's ACE framework creates game characters that react to player behavior in real time.
Wooldridge-Jennings mapping: Autonomy — plays without human guidance. Reactivity — evaluates opponent moves and board state changes within strict time controls. Proactiveness — develops multi-move strategies and sacrifices short-term material for positional advantage. Social ability — plays against human and artificial opponents, adapting to opponent style.
Infrastructure: Training required 5,000 TPUs for AlphaGo Zero. Inference runs on commodity GPUs but needs deterministic latency for competition play. What breaks: GPU memory exhaustion during deep search truncates the evaluation tree, causing blunders at critical positions. Non-deterministic compute scheduling produces inconsistent play strength — an agent that plays at 3200 Elo one game and 2800 the next is unreliable for any serious deployment.
3. Industrial Control Systems
ReactiveSiemens MindSphere monitors over 1.5 million connected devices across manufacturing plants worldwide. ABB's Ability platform manages power grids, industrial robots, and process control systems. These are not the thermostats of your textbook — they are intelligent agents perceiving thousands of sensor streams simultaneously, maintaining physics-based models of industrial processes, and making real-time adjustments to optimize output, energy efficiency, and equipment lifespan. A modern paper mill runs 800+ control loops, each one an intelligent sub-agent coordinating with the others.
Wooldridge-Jennings mapping: Autonomy — operates continuously, adjusting setpoints without operator input. Reactivity — responds to sensor anomalies within milliseconds. Proactiveness — predicts equipment degradation and schedules preemptive maintenance. Social ability — coordinates with other control agents across the plant, communicates status to human operators and SCADA systems.
Infrastructure: Edge compute nodes for low-latency local control, with cloud aggregation for fleet-wide optimization. Redundant PLCs with sub-10ms failover. What breaks: Network partition between edge and cloud causes agents to operate on stale global models. Sensor drift without calibration feedback produces gradual quality degradation that is invisible until product defects appear downstream. The 2021 Oldsmar water treatment hack demonstrated what happens when an intelligent control agent's actuators get compromised — the system tried to raise sodium hydroxide to lethal levels.
Category 2 — Deliberative Intelligence
Agents That Think Before They Act
Deliberative intelligent agents maintain explicit world models, formulate multi-step plans, and reason about future states before committing to action. The latency tolerance here is higher — seconds to hours instead of milliseconds — but the compute requirements are often larger because planning is expensive. The infrastructure challenge is sustained cognitive load: these agents run long reasoning chains that can exhaust memory, accumulate context errors, and enter infinite loops when the planning gets stuck.
4. Research Agents — Perplexity Deep Research
DeliberativePerplexity's Deep Research mode does not just search the web. It decomposes a research question into sub-queries, executes parallel web searches, synthesizes intermediate findings, identifies knowledge gaps, formulates follow-up queries, and produces a structured research report with citations. A single Deep Research query can trigger 30-50 separate web searches, read and analyze hundreds of pages, and produce a 2,000-word synthesis in under five minutes. By March 2026, Perplexity reports processing over 100 million queries per week across its platform.
Wooldridge-Jennings mapping: Autonomy — once initiated, the research plan executes without user intervention. Reactivity — adjusts the research plan based on what intermediate searches find (or fail to find). Proactiveness — identifies knowledge gaps and generates sub-queries the user never asked for. Social ability — interacts with web servers, APIs, and citation databases as information-providing agents.
Infrastructure: High-memory servers for parallel search execution, large context windows for synthesis, and persistent state for multi-minute research sessions. What breaks: Context window overflow during long research chains causes the agent to “forget” earlier findings. Network timeouts on source websites produce incomplete research with false confidence. Without persistent session state, a crash at minute four of a five-minute research task wastes all accumulated work.
5. Coding Agents — Devin, Claude Code, SWE-Agent
DeliberativeDevin launched in March 2024 as the first “AI software engineer” with a dedicated cloud sandbox for each coding session. By 2026, the category has expanded: Claude Code operates as an agentic terminal with file system access and shell execution, SWE-Agent achieves 23% autonomous resolution on the SWE-bench benchmark, and GitHub Copilot Workspace plans multi-file edits from natural language specifications. These are not code completion tools. They are deliberative agents that decompose programming tasks into sub-goals, generate implementation plans, write code, run tests, observe failures, and iterate until the tests pass.
Wooldridge-Jennings mapping: Autonomy — executes multi-step coding tasks without human intervention (Devin sessions average 15-30 minutes). Reactivity — observes test output, error messages, and build logs to adjust implementation. Proactiveness — generates test cases before writing code, refactors proactively when detecting code smells. Social ability — interacts with version control, CI/CD pipelines, and code review platforms.
Infrastructure: Each session requires a sandboxed compute environment with shell access, file system, network for package installation, and 8-32 GB RAM. What breaks: Memory leaks during long sessions cause OOM kills mid-task. Without watchdog auto-restart, a crashed coding agent leaves half-written code in a broken state. Non-isolated environments allow agent file operations to affect host system stability.
6. Supply Chain Optimizers
DeliberativeAmazon's supply chain AI manages inventory positioning across 175+ fulfillment centers in North America alone, deciding what to stock where based on demand forecasts, shipping costs, delivery time commitments, and seasonal patterns. Flexport's intelligent agents optimize container routing across global shipping lanes, negotiating between 20+ variables per shipment decision. These agents think in horizons of days to weeks, maintaining complex world models of global logistics that include weather patterns, port congestion, geopolitical risks, and currency fluctuations.
Wooldridge-Jennings mapping: Autonomy — repositions inventory and reroutes shipments without human approval for decisions below threshold value. Reactivity — adjusts plans when port closures, weather events, or demand spikes are detected. Proactiveness — pre-positions inventory for predicted demand events weeks in advance. Social ability — coordinates with supplier agents, carrier APIs, customs systems, and warehouse management systems.
Infrastructure: High-memory compute for optimization solvers, persistent databases for world model state, and reliable API connectivity to dozens of external logistics systems. What breaks: Stale demand data produces optimal solutions to yesterday's problem. API timeouts with carrier systems cause the agent to plan routes that do not exist. The 2021 Suez Canal blockage showed what happens when a supply chain agent's world model encounters a state it was never trained on — most systems required manual human override for weeks.
Category 3 — Social Intelligence
Agents That Think Together
Social ability is the most underestimated of the four properties. The textbooks treat it as an afterthought — “the agent can communicate with other agents.” In production, social intelligence is the property that determines whether a system of agents produces emergent capability or emergent chaos. The production AI agent landscape is moving from single-agent architectures to multi-agent systems, and the coordination problem is where most deployments fail.
7. Multi-Agent Debate Systems
SocialBoth Anthropic and Google DeepMind published research in 2025 showing that multiple LLM agents debating a question produce more accurate answers than a single agent reasoning alone. The architecture: one agent generates an initial response, a second agent critiques it, the first agent revises, and the process repeats for 3-5 rounds. Google's Society of Mind framework extends this to specialized agents — a fact-checker, a logic validator, a creative thinker — each contributing from its area of strength. Production deployments at consulting firms use debate architectures for investment analysis, generating bull and bear cases simultaneously.
Wooldridge-Jennings mapping: Autonomy — each debating agent operates independently with its own reasoning process. Reactivity — each agent perceives and responds to the other agents' arguments. Proactiveness — agents actively seek weaknesses in opposing arguments and construct counterexamples. Social ability — the defining property: structured argumentation protocols between agents.
Infrastructure: Multiple concurrent LLM inference sessions with message passing, shared context state, and process isolation between agents. What breaks: Without process isolation, one agent's memory leak crashes all debaters. Message ordering bugs cause agents to respond to stale arguments. The debate can enter infinite agreement loops where agents converge on a confidently wrong consensus — infrastructure monitoring of token consumption and argument diversity is essential.
8. Customer Escalation Chains
SocialKlarna's AI customer service system handles 2.3 million conversations per month, resolving 67% of inquiries without human involvement. But the interesting engineering is not in the tier-1 resolution — it is in the escalation architecture. When the frontline agent detects a query it cannot handle (billing dispute, regulatory complaint, emotionally distressed customer), it does not just hand off. It transfers a structured context package to a specialized tier-2 agent: conversation history, customer sentiment score, identified issue category, attempted resolutions, and confidence level. The tier-2 agent might be another AI with domain-specific training, or a human agent receiving the AI-prepared brief.
Wooldridge-Jennings mapping: Autonomy — handles the full customer interaction lifecycle without human scheduling. Reactivity — detects customer sentiment shifts and topic changes in real time. Proactiveness — identifies issues the customer has not explicitly raised based on account history. Social ability — maintains structured communication with tier-2 agents, human operators, CRM systems, and payment processors.
Infrastructure: Always-on servers with vector databases for knowledge retrieval, WebSocket connections for real-time chat, and reliable inter-agent message queuing. What breaks: Context loss during escalation handoffs produces the “can you repeat your problem?” experience that destroys customer trust. Agent downtime during peak hours (typically 2x normal query volume) causes cascading queue buildup. Without self-healing restart, a crashed escalation agent creates a black hole where transferred conversations vanish.
9. Collaborative Coding Teams
SocialThe single coding agent (example 5) is evolving into multi-agent coding teams. Microsoft's AutoGen framework and CrewAI enable architectures where separate agents handle planning, implementation, testing, and code review. One agent writes the implementation, another writes tests against the spec, a third runs the tests and reports failures back to the implementer, and a fourth reviews the final code for style, security, and architectural consistency. Early production deployments at companies like Factory AI and Poolside report 20-40% improvement in first-pass code quality compared to single-agent approaches.
Wooldridge-Jennings mapping: Autonomy — each agent in the team operates independently within its role. Reactivity — test agent reacts to code changes, review agent reacts to completed implementations. Proactiveness — planning agent decomposes tasks, implementation agent anticipates edge cases. Social ability — structured message passing between agents with role-specific protocols.
Infrastructure: Isolated compute environments per agent, shared file system access with conflict resolution, and orchestration layer for task routing. What breaks: File system race conditions when two agents edit the same file. Deadlocks when the implementer waits for test results while the tester waits for implementation. Without resource isolation, one agent's runaway process starves the others, causing the whole team to stall.
Category 4 — Learning Intelligence
Agents That Get Better Over Time
Learning intelligence is the property that separates a static deployment from a living system. These agents do not just operate in an environment — they accumulate experience, update their models, and genuinely improve their performance over time without explicit reprogramming. The infrastructure challenge is persistence: a learning agent that loses its accumulated state is lobotomized back to its initial deployment quality.
10. Recommendation Engines That Improve
LearningNetflix's recommendation system processes behavioral signals from 260 million subscribers across 190 countries, and it is not a static model serving cached predictions. It is a learning intelligent agent that updates user embeddings in near real-time based on viewing behavior, browse patterns, search queries, and even hover duration over thumbnails. Spotify's Discover Weekly generates 30 billion personalized track recommendations per week using a similar architecture. TikTok's For You page adapts within minutes of observing new user behavior, demonstrating learning latency that was unthinkable five years ago.
Wooldridge-Jennings mapping: Autonomy — operates continuously, selecting content without human curation. Reactivity — responds to user actions (skip, watch, rewatch, search) within seconds. Proactiveness — surfaces content the user has not searched for, anticipating preferences before they are expressed. Social ability — incorporates collaborative filtering from millions of other users as implicit social signals.
Infrastructure: Feature stores for real-time user embeddings, ML serving infrastructure for sub-100ms inference, and massive persistent storage for behavioral logs. What breaks: Feature store corruption causes recommendations to reset to cold-start quality for affected users. Training pipeline failures mean the agent stops learning while the world changes around it — it starts recommending content from cultural moments that have passed. Netflix reported that a 2019 recommendation system outage cost an estimated $1 million per hour in engagement-driven revenue.
11. Adaptive Fraud Detection
LearningStripe Radar processes billions of transactions per year, and the critical word is “adaptive.” Static fraud rules catch yesterday's fraud patterns. Adaptive fraud agents learn from every transaction outcome — approved, declined, charged back, flagged by the cardholder — and update their risk models in near real-time. Featurespace's ARIC platform reports catching fraud patterns within 24 hours of emergence, before any rule could be written. PayPal's fraud detection system evaluates 10+ million transactions daily with models that retrain on a continuous feedback loop of human reviewer decisions.
Wooldridge-Jennings mapping: Autonomy — blocks or approves transactions in real-time without human review for 95%+ of volume. Reactivity — responds to individual transaction signals (amount, location, device, velocity) within milliseconds. Proactiveness — identifies emerging fraud patterns before they are reported, adjusts risk thresholds preemptively during detected attack campaigns. Social ability — communicates with issuing banks, merchant systems, and human fraud analysts through structured alert protocols.
Infrastructure: Real-time feature computation, sub-50ms inference latency, and continuous model retraining pipelines with A/B testing for new model versions. What breaks: Training data poisoning — if fraudsters learn to generate transactions that look legitimate to the feedback loop, the agent learns the wrong lessons. Retraining pipeline failures freeze the model in a past state while fraud tactics evolve. Latency spikes during peak shopping periods force the agent to either approve blindly or block legitimate transactions, both of which erode the system.
12. Self-Tuning Infrastructure Agents
LearningGoogle's Autopilot manages resource allocation across Borg clusters serving billions of requests. Amazon's auto-scaling agents adjust EC2 fleet sizes based on learned traffic patterns, not just threshold rules. Kubernetes Vertical Pod Autoscaler uses historical resource consumption data to right-size container requests. These are infrastructure agents that manage infrastructure — a recursive property that makes them fascinating from a Wooldridge-Jennings perspective. They learn the behavior of the workloads they manage and proactively adjust resource allocation to optimize cost, performance, and reliability simultaneously.
Wooldridge-Jennings mapping: Autonomy — adjusts resource allocation without human approval within configured bounds. Reactivity — detects traffic spikes and resource pressure in real-time. Proactiveness — pre-scales capacity before predicted demand events (learned from historical patterns). Social ability — coordinates with scheduling systems, monitoring agents, and capacity planning tools.
Infrastructure: Time-series databases for historical metrics, ML inference for predictive scaling, and reliable control plane access for resource adjustment. What breaks: The recursive problem: the infrastructure agent needs infrastructure. If the monitoring system that feeds the agent goes down, the agent becomes blind. If the agent's own compute gets scaled down by another agent, it cannot scale anything. Google's internal post-mortems document cases where auto-scaling agents created oscillation patterns — scaling up triggered alerts that triggered scaling down that triggered scaling up — consuming resources in a feedback loop that served no workload.
The Pattern Across All 12 Examples
Something becomes obvious when you lay these 12 examples side by side. Every intelligent agent, regardless of category, requires the same four infrastructure capabilities mapped directly to the four Wooldridge-Jennings properties. This is not a coincidence. It is a consequence of what intelligence means in an operational context.
Autonomy requires process supervision
Every example above fails the autonomy test when the process crashes and stays down. HFT bots need sub-microsecond failover. Coding agents need watchdog restart with state preservation. Customer escalation chains need zero-downtime deployments. The specific mechanism varies but the requirement is universal.
Reactivity requires low-latency observability
An agent that cannot perceive its environment cannot react to it. Nanosecond market data feeds for HFT. Real-time sensor streams for industrial control. WebSocket connections for customer support. Token consumption monitoring for debate systems. The specific sensors vary but the requirement for timely perception is constant.
Proactiveness requires persistent state and scheduling
Goal-directed behavior requires memory of goals. Research agents need session state for multi-minute plans. Supply chain optimizers need world models that persist across decision cycles. Recommendation engines need learned user preferences that survive restarts. Without persistence, every session starts from zero and proactiveness is impossible.
Social ability requires reliable networking and isolation
Multi-agent debate needs message passing without message loss. Customer escalation needs context transfer without corruption. Collaborative coding needs file system coordination without race conditions. Social ability is only as reliable as the communication channel, and the communication channel is an infrastructure problem.
Strip away any one of these infrastructure capabilities and the corresponding property degrades. Strip away two and you no longer have an intelligent agent — you have an expensive API wrapper with pretensions. This is the thesis of this article distilled to its core: the type of agent you are building matters less than the infrastructure substrate you are building it on. Intelligence without infrastructure is just a hallucinating API call.
Infrastructure Requirements by Intelligence Type
| Intelligence Type | Latency | RAM | Uptime | Critical Need |
|---|---|---|---|---|
| Reactive | µs–ms | 16–64 GB | 99.999% | Failover speed |
| Deliberative | s–min | 8–64 GB | Per-session | State persistence |
| Social | ms–s | 16–32 GB × N agents | 24/7 | Process isolation |
| Learning | ms–hr | 32–128 GB | 24/7 | Durable storage |
The Infrastructure That Makes Intelligence Possible
I want to be direct about how osModa connects to these 12 examples. We do not make your agent smarter. We do not improve your model's reasoning or reduce hallucination rates. What we do is provide the environmental substrate that preserves the four Wooldridge-Jennings properties in production — so that the intelligence you built in the lab actually survives contact with the real world.
Rust-based watchdog for autonomy: Our self-healing watchdog does not just check “is the process alive.” It monitors functional health — is the agent responding coherently, or has it entered a hallucination loop? When an agent degrades, the watchdog detects the anomaly and restarts within 6 seconds with preserved state. That is how you keep deliberative agents like coding agents and research agents autonomously operational through multi-hour sessions.
NixOS atomic rollback for rational correction: When an agent's actions lead to a degraded environment state — a dependency that breaks, a config change that corrupts performance — SafeSwitch reverts the entire system to the last known-good state at the platform level. Not container-level. Full system state: binaries, configuration, dependencies, environment variables. The supply chain optimizer that updates its world model incorrectly gets a clean do-over.
SHA-256 audit chain for accountable agency: Every intelligent agent example above acts in the world with real consequences. The tamper-evident hash chain records every action, every state transition, every environmental interaction. When a fraud detection agent makes a blocking decision that needs to be explained to a regulator, the audit trail provides complete, verifiable history.
Dedicated resources for predictable performance: No shared tenancy. Intelligent agents have unpredictable resource consumption — a simple query might use 2 GB while a deep reasoning chain uses 16 GB. Shared infrastructure means noisy neighbors, memory contention, and the kind of latency spikes that destroy reactivity. Dedicated Hetzner servers starting at $14.99/month, with resource isolation that guarantees your agent's performance depends only on its own behavior.
The Thesis, Revisited
I started with a provocation: most intelligent agent content online is stuck in 2005. After walking through 12 production systems, I think the more precise statement is that the textbooks got the definition right and the examples wrong. The four Wooldridge-Jennings properties are as valid today as they were in 1995. What has changed is the infrastructure required to sustain those properties in production environments where failure costs real money.
A high-frequency trading bot that crashes every three hours is not autonomous. A research agent that loses its accumulated findings on restart is not proactive. A customer escalation chain that drops context during handoffs is not socially able. A fraud detection system that stops learning is not reactive to evolving threats. In every case, the failure is not in the algorithm. It is in the infrastructure.
Intelligence without infrastructure is just a hallucinating API call. The examples in this article prove it. The question for anyone building intelligent agents in 2026 is not “which model should I use?” The models are good enough. The question is: does your infrastructure preserve the four properties that make your agent genuinely intelligent when the real world starts happening to it?
Frequently Asked Questions
What is an intelligent agent in artificial intelligence with examples?
An intelligent agent is a system that exhibits four properties formalized by Wooldridge and Jennings in 1995: autonomy (operates without direct human control), reactivity (perceives and responds to environmental changes), proactiveness (takes initiative toward goals), and social ability (interacts with other agents). Real examples include AlphaGo (reactive intelligence — evaluating 200 million board positions per second), Perplexity Deep Research (deliberative intelligence — decomposing research queries into multi-step search plans), and Netflix's recommendation engine (learning intelligence — continuously adapting to 260 million user behavior profiles). The key distinction from a simple program is the closed feedback loop: an intelligent agent modifies its behavior based on environmental outcomes.
What is the difference between an AI agent and an intelligent agent?
An AI agent is any software system that uses artificial intelligence to perceive and act in an environment. An intelligent agent is a stricter classification requiring all four Wooldridge-Jennings properties operating simultaneously: autonomy, reactivity, proactiveness, and social ability. A chatbot answering customer questions is an AI agent — it perceives input and generates responses. But if it cannot escalate to human agents (social ability), proactively identify emerging issues (proactiveness), or operate without scheduled restarts (autonomy), it fails the intelligent agent test. Most production AI agents in 2026 satisfy two or three properties. Systems satisfying all four are rarer and require significantly more sophisticated infrastructure.
What are the four properties of an intelligent agent?
The four properties, formalized by Wooldridge and Jennings (1995), are: (1) Autonomy — the agent operates without direct human intervention and controls its own actions and internal state. (2) Reactivity — the agent perceives its environment through sensors and responds to changes in a timely fashion. (3) Proactiveness — the agent takes initiative, pursuing goals through goal-directed behavior rather than only reacting to stimuli. (4) Social ability — the agent interacts with other agents, whether human or artificial, through communication protocols. Each property imposes specific infrastructure requirements: autonomy needs process supervision and crash recovery, reactivity needs low-latency I/O, proactiveness needs persistent scheduling, and social ability needs reliable networking and protocol support.
What is a reactive intelligent agent example?
High-frequency trading bots at firms like Citadel Securities and Jane Street are canonical reactive intelligent agents. They perceive market data streams at microsecond resolution, react to price movements and order book changes within nanoseconds, and execute trades without human intervention. Citadel Securities alone processes roughly 27% of U.S. equity volume. What makes them genuinely intelligent rather than simple reflex agents is internal state: they maintain models of market microstructure, track their own position risk, and adjust reaction thresholds based on volatility regimes. The infrastructure requirement is extreme — co-located servers within exchange data centers, FPGA-based execution, and sub-microsecond failover. A 1-millisecond delay can cost millions.
What is the difference between reactive and deliberative intelligent agents?
Reactive intelligent agents respond to environmental stimuli in near real-time using pattern matching and learned reflexes — they perceive and act quickly but do not engage in long-horizon planning. High-frequency trading bots and game AI like AlphaGo's move evaluation are reactive. Deliberative intelligent agents maintain explicit world models, formulate plans, and reason about future states before acting. Research agents like Perplexity Deep Research and coding agents like Devin are deliberative — they decompose complex tasks into sub-goals, generate multi-step plans, and revise strategies based on intermediate results. In practice, most production intelligent agents combine both: reactive components for real-time response and deliberative components for strategic planning.
Can an intelligent agent work with other agents?
Social ability is one of the four core properties of intelligent agents, and multi-agent collaboration is increasingly common in production. Examples include multi-agent debate systems where separate LLM agents argue opposing positions to improve reasoning quality (used by Anthropic and Google DeepMind in research), customer escalation chains where tier-1 agents hand off to specialized tier-2 agents with full context transfer, and collaborative coding teams where separate agents handle planning, implementation, testing, and code review. The infrastructure challenge is coordination: agents need reliable message passing, shared state management, and conflict resolution protocols. Without robust networking and process isolation, multi-agent systems degrade into confused single agents talking to themselves.
What infrastructure does an intelligent agent need?
Each of the four Wooldridge-Jennings properties maps to specific infrastructure requirements. Autonomy requires process supervision (watchdogs that detect crashes and restart within seconds), persistent state storage, and resource isolation. Reactivity requires low-latency I/O, event-driven architectures, and real-time monitoring. Proactiveness requires persistent scheduling, goal state management, and planning compute. Social ability requires reliable networking, message queuing, and protocol support for agent-to-agent communication. The non-obvious requirement is infrastructure observability — an intelligent agent that cannot detect its own degradation loses the feedback loop that makes it intelligent. osModa provides this substrate through NixOS-based servers with Rust watchdogs, atomic rollback, and SHA-256 audit chains.
Why do intelligent agents fail in production?
The dominant failure mode is infrastructure-induced property loss. An intelligent agent deployed on unreliable infrastructure progressively loses its defining properties. Without crash recovery, it loses autonomy — a crashed agent is not autonomous. Without low-latency monitoring, it loses reactivity — an agent that cannot perceive environmental changes cannot react to them. Without persistent scheduling, it loses proactiveness — an agent that only responds to explicit triggers is not taking initiative. Without reliable networking, it loses social ability — an agent that cannot communicate with other agents is socially isolated. When enough properties degrade, what remains is not an intelligent agent but an expensive API wrapper that occasionally produces useful output. This is why the thesis of this article is that intelligence without infrastructure is just a hallucinating API call.