Agent Systems Group Guide
Group: Agent Systems | Previous group exit: define and safely execute one controlled action | This group exit: build a stateful, recoverable, permission-bounded autonomous loop
1. Overview
BLUF: this group upgrades a single action into a sustained loop: the model observes, decides, and acts until a stop condition fires. Autonomy brings new failure modes — state bloat, runaway loops, error propagation, unsupervised side effects. Ten pages run from mental model, through patterns, to control plane and scale.
| Order | Page | Question it answers |
|---|---|---|
| Front door | Mental model & runtime | What composes the loop? When is an agent worth it? |
| Patterns | Design patterns | ReAct, routing, planner-executor, reflection — which and when? |
| State | State & memory | Where do cross-step / cross-turn data live, and how to recover? |
| Control plane | Hooks | Lifecycle interception and automated policy gates |
| Control plane | Recovery & HITL | Which state do you return to? When must a human decide? |
| Environment fit | Computer use | observe→act→verify when only a visual UI exists |
| Deterministic alternative | Workflow patterns | When not to use an agent |
| Capability packaging | Skills · Plugins | Distributing procedural knowledge and capability packages |
| Scale | Subagent / multi-agent | Delegation and orchestration within one trust domain |
Boundary: delegation across processes, organizations, or trust domains is not here — go through Interoperability (A2A et al.) first.
2. Usage
- First contact with agents: read top-down starting from Mental model & runtime.
- Already running agents in production: jump to Recovery & HITL and Hooks.
- Unsure an agent is warranted: read the reverse decision in Workflow patterns first.
3. Principles
Three loop invariants: stop conditions must be explicit (steps / tokens / wall-clock budget), state must be checkpointable, side effects must pass Action-group execution discipline. Missing any one, autonomy only amplifies failure.
4. Development
Pages share one debugging grammar: symptom → evidence (trace / checkpoint) → fix → done criteria. The multi-agent debug unit is the delegation; the single-agent unit is the loop step.
5. Resource Library
Each topic page carries its own four-tier route; group-level entries are Anthropic's Building Effective Agents and the OpenAI Agents guide (see agent-runtime resource tables).