Agentic Workflows

Patterns for turning a one-shot API call into something that does things — iterates, remembers, coordinates. This section extends the Tool Use loop into a full agent architecture.

Start here

• Loops — the plan / act / observe cycle that turns tool use into an agent.

Going further

• Memory — what to do when tool outputs start eating your context window.
• Multi-agent — when splitting work across specialized agents helps vs. hurts.

The progression

Each page adds one capability to the previous one:

flowchart LR
    L[Tool-use loop] --> A[Agent with iteration cap]
    A --> M[Agent with memory]
    M --> X[Multiple cooperating agents]

The rule of thumb is to stop as soon as the current layer is enough for your task. Extra layers multiply cost, latency, and debugging pain.

Further reading

• Building Effective Agents (Anthropic) — taxonomy of agent patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) with clear guidance on when to use each. Maps closely to the patterns in this section.