agency-multi-agent-systems-architect

Start by selecting a topology (sequential, parallel, hierarchical, or evaluator-optimizer), then define each agent's input/output contract, fallback chain, and permission scope. Add structured trace logging with a shared trace_id and eval suites before deployment.

Default to hierarchical orchestrator-subagent for complex dynamic tasks, sequential chains for linear workflows, and parallel fan-out for independent subtasks. Avoid mesh topologies unless negotiation is required, since they are hardest to debug.

Classify failures as hard, silent, partial, contradiction, cascade, loop, or context failures, then apply circuit breakers and a fallback chain: primary agent, narrowed fallback, degraded rule-based output, then human escalation. The system must always produce a structured response.

Place blocking approval gates on irreversible actions, high blast-radius operations, low-confidence outputs below 0.7, and regulatory-exposure cases. Use advisory flags or sampling gates for reversible, high-volume actions to avoid over-escalation.

Isolate external content from system instructions, use a sanitizer agent to extract structured data from untrusted input, and enforce schema validation on outputs. Flag any agent output containing instruction-like language for quarantine.

Mesh networks carry the highest complexity: circular dependencies, consensus deadlock, exponential context growth, and difficult debugging. They require a moderator agent, explicit consensus mechanism, and a circuit breaker that escalates to a human after N rounds.