Agent Architecture & Orchestration
One agent doing everything is a single point of failure with a full context window.
What This Covers
Section titled “What This Covers”How multi-agent work gets structured: decomposing a goal into agent-sized tasks, choosing an orchestration topology, validating outputs at the seams, and recovering when an agent fails mid-flight.
Key Principles
Section titled “Key Principles”- Decompose work into focused, single-purpose agents
- Validate outputs before acting on them
- Build recovery into the architecture, not as an afterthought
- Failures are the most expensive tokens; design to minimize them
Orchestration Topologies
Section titled “Orchestration Topologies”- Fanout: independent subtasks run in parallel, results merged (Subagent Fanout)
- Pipeline: staged handoffs where each stage’s output is the next stage’s input (Pipeline Orchestration)
- Adversarial: one agent produces, another attacks the result (Adversarial Review)
- Consensus: several models answer independently, a judge reconciles (Multi-Model Consensus)
Key Considerations
Section titled “Key Considerations”Task Boundaries
Section titled “Task Boundaries”- Can the subtask be briefed in one message, with no mid-flight clarification?
- Does it share mutable state (a branch, a file) with a concurrent agent? Disjoint scopes or serialize.
- Is the acceptance criterion checkable by the orchestrator without redoing the work?
Trust and Validation
Section titled “Trust and Validation”- An agent’s report is a claim, not evidence: verify the artifact, not the summary
- Fabrication happens; validation at the seam is what keeps it out of deliverables
- Use Verification Loops at every handoff that feeds a downstream action
Recovery
Section titled “Recovery”- What happens when an agent times out, drifts from its brief, or returns garbage?
- Prefer retriable, idempotent subtasks; design so a failed agent can be relaunched cold
- Keep the orchestrator lean: coordination and decisions, not implementation
Anti-patterns
Section titled “Anti-patterns”- One mega-agent with an ever-growing prompt instead of decomposition
- Trusting subagent reports without inspecting outputs
- Parallel agents sharing a branch or file and merging by hope
- Orchestrator doing the work itself, burning the coordination context