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
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