01

First prove that you need a team

The most advanced architecture is not automatically the best. A deterministic workflow fits known steps; one agent with good tools preserves context and simplifies evaluation. Split the system when observed conditional complexity, tool overlap or distinct responsibilities degrade the result.

Fixed workflow

Predictable steps, explicit rules and programmatic gates between stages.

Single agent

The route varies, but one responsibility and a coherent tool set are enough.

Agent team

Domains, permissions, context or criteria genuinely differ and benefit from isolation.

Do not build

The decision cannot be evaluated, failure cost is unacceptable or conventional code solves the work better.

02

Choose where control lives

OpenAI describes two general patterns: a manager that invokes specialists as tools and a decentralized system where agents transfer execution through handoffs. A practical third pattern, orchestrator-workers, decomposes variable work and synthesizes the results.

Manager

One voice retains control, delegates parts and composes the final answer. Useful when the user needs one accountable owner.

Handoffs

A specialist takes over execution and recent context. Useful when full responsibility changes hands.

Orchestrator-workers

The coordinator creates case-specific subtasks, allows safe parallelism and assembles artifacts.

Deterministic code

Critical transitions, permissions, limits and irreversible effects do not need model judgment.

03

Issue work packets, not wishes

A specialist should not receive “solve this.” Its task contract should support action and demonstrate why the result is acceptable.

  • Objective and observable outcome.
  • Authorized inputs and source versions.
  • Allowed tools and explicitly forbidden actions.
  • Output schema and retained artifacts.
  • Acceptance tests and required evidence.
  • Maximum turns, time, cost and retries.
  • Escalation rule when human judgment is missing.
04

Treat tools as boundaries

MCP standardizes connections between a host, isolated clients and servers exposing capabilities. Composition does not remove security: the host retains authorization, consent and context; capabilities are negotiated and tools declare schemas and effects.

Least privilege

Each role sees only the sources and actions required by its assignment.

Explicit schemas

Arguments, outputs and errors must be validated before proceeding.

Visible effects

Sending, deleting, publishing or paying needs approval proportional to impact.

Idempotency

A retry must not duplicate actions or corrupt state.

Useful records

Keep decisions, tools, results and errors without retaining unnecessary secrets.

05

Separate construction from verification

The result must pass controls that do not depend on trusting the producing agent's prose. Evaluation can combine deterministic checks, human review and model graders calibrated against human-reviewed examples.

  • Validate the schema before judging content.
  • Run builds, tests or queries against the actual artifact.
  • Assess the final answer as well as correct intermediate results.
  • Turn every known failure into a permanent regression case.
  • Compare quality, cost, latency, calls, turns and retries.
  • Reserve human judgment for usefulness, risk and subjective criteria.
06

Use four exit states

An agent loop needs observable stop conditions. This operational synthesis prevents “I am done” from remaining a model opinion.

PASS

The artifact exists, required tests pass and evidence is tied to the evaluated version.

REVISE

The failure is reproducible, an owner is known and budget remains for a bounded correction.

BLOCKED

A dependency, permission or external source is missing; the exact cause is retained without inventing a result.

ESCALATE

A business decision, risk authorization or subjective human criterion is required.

07

A reusable definition of done

Fill this list before work starts. Then require the coordinator to return it with evidence instead of an optimistic summary.

  • Exact artifact and version identified.
  • Numbered acceptance criteria.
  • Tests executed with result and date.
  • External effects approved by the right authority.
  • Cost, duration, calls and retries recorded.
  • Limitations and uncovered cases declared.
  • Final PASS, REVISE, BLOCKED or ESCALATE state with rationale.