Appendix - Agent Patterns & Multi-Agent
What We Learned
- Use chaining, routing or parallelism when the steps are already predictable.
- Start with one agent and good tools; add specialization when measured gains justify it.
- Choose orchestrator-workers, handoffs or other topologies around ownership and dependencies.
- Give agents concise task contracts; control shared state, budgets and recovery.
- Measure success, latency and cost against a simpler baseline; verify results externally.
Key Acronyms, Concepts & Jargon
| Term | Short meaning |
|---|---|
| Chaining / routing | Sequential dependent steps / selecting a path for a request. |
| Fan-out / fan-in | Dispatching independent work / collecting and combining its results. |
| Orchestrator-workers | A coordinator delegates scoped tasks and integrates worker results. |
| Evaluator-optimizer | An evaluator supplies feedback to an iterative improvement step. |
| Sub-agent / handoff | Delegated worker under a parent / transfer of control to another agent. |
| Topology | Arrangement of agents and the communication paths between them. |
| Supervisor / peer network | Central coordinator / agents coordinating directly with one another. |
| Task contract | Goal, inputs, constraints and expected output for delegated work. |
| Shared state / ledger | Common mutable information / recorded actions and ownership. |
| Idempotency | Retrying an operation preserves the intended effect of one execution. |
| MAST | Multi-Agent System Failure Taxonomy: categories for diagnosing coordination failures. |
| Ablation / baseline | Remove a component to measure its value / reference system for comparison. |
| Context isolation | Keeps unrelated worker details out of other agents' contexts. |