07 section

Agentic Systems

Building systems that take actions rather than answer questions: reasoning loops, tool protocols, multi-agent teams, planning, failure recovery, human oversight, sandboxing and trajectory evaluation

10 pages 32 min 6,627 words

01 the section

What is in here

The first four pages build a working agent, the next three keep it from going wrong, and the last three decide whether it can ship — human oversight, blast-radius control, and knowing whether it actually works. The memory page here is the agent's view; Memory and State covers the storage architecture behind it. Numeric order is the right order.

01 3 min Agent Fundamentals What separates an agent from a chatbot — reasoning model, tools, memory, environment feedback — and the autonomy levels you should deliberately choose between agentsfundamentalsreasoning 02 3 min Reasoning Loops: ReAct and Beyond Control flow for agents, from ReAct's thought-action interleave to Reflexion retries, Plan-and-Solve, and explicit graph-shaped flow engineering when loops stop being enough agentsreasoningorchestration 03 5 min Tool Use and MCP Tool schemas, the Model Context Protocol with its Streamable HTTP and auth updates, computer-use tools, and how MCP differs from plain function calling tool-useagentsstructured-output 04 3 min Multi-Agent Orchestration One agent with fifty tools versus a team of specialists: supervisor hierarchies, pipelines, peer-to-peer swarms, and where shared state lives between them orchestrationagentsstate 05 3 min Agent Memory and State The three memory tiers an agent needs — working context, episodic trajectories, semantic profile — and what each one is actually stored in memorystateagents 06 3 min Planning and Decomposition Stopping an agent from wandering: linear versus hierarchical plans, static versus replanned execution, recursive task decomposition, and tree search over candidate action paths planningagentsreasoning 07 3 min Error Handling and Recovery Agents fail as hallucinated tools, schema violations, dead APIs and infinite loops; each gets a detection signal, a recovery path, and a checkpoint to roll back to reliabilityagentsstate 08 3 min Human-in-the-Loop Patterns Where to put the human: approval gates, interrupts and breakpoints, time-travel state editing, shared scratchpads, and escalation triggered by the model's own confidence oversightagentsreliability 09 3 min Agentic Security and Sandboxing An injected agent does not leak data, it acts — so isolate execution in sandboxes, scope credentials to minimum agency, proxy every call, and log the trail securityagentsoversight 10 3 min Evaluating Agentic Systems Scoring the whole trajectory rather than the final answer: task-completion benchmarks, cost and step-count metrics, LLM judges for step quality, and production A/B design evaluationagentsproduction