Skip to content

Multi-Agent Pattern Decision Record (PDR)

Use this template to document a multi-agent collaboration pattern decision. The goal is not to prove that the design is “advanced,” but to explain why it fits the current constraints better than a simpler baseline and when to fall back.

1. Basic information

Field Content
PDR ID PDR-000
Decision name
Status Proposed / Accepted / Superseded
Date
Decision makers
Related ADR / requirement / incident

2. Task and constraints

Objective

Why consider multiple agents

Candidate boundary Present? Evidence
Domain boundary
Data boundary
Tool boundary
Permission boundary
State boundary
Failure boundary
Team boundary

Key constraints

  • Latency / SLO:
  • Cost budget:
  • Data and permissions:
  • Risk and approval:
  • Availability and degradation:
  • Evidence and audit:

3. Control model

Question Decision
Who decides the next step?
Who owns the global objective and outcome?
How are the agents connected?
What is the actual boundary of each agent?
How does the system complete, partially complete, or stop?

4. Candidate options

Option Benefits Risks / costs Advance to validation?
Deterministic workflow
Single agent
Handoff
Supervisor
Router + Fan-Out/Join
Other

5. Decision

Selected pattern and composition

Why not choose a simpler baseline

Ownership and state writers

State / artifact Owner Concurrency and versioning strategy
Global Task
Agent Local State
Delegation Ledger
Artifact / Evidence
Policy / Approval

Collaboration contracts

  • Delegation Contract:
  • Result Contract:
  • Evidence Contract:
  • Join Contract:

Permission and risk boundaries

  • Effective permissions after delegation:
  • Whether further delegation is allowed:
  • Maximum risk level for automatic execution:
  • Human approval points:

6. Failure and convergence

Scenario Handling
Single-agent timeout
Invalid Result Schema
Conflicting evidence
Dependency unavailable
Duplicate message / result
Budget exhausted
Repeated lack of progress
Unapproved high-risk action

7. Validation plan

Metric Simple baseline Target Observation window
Task success rate
Result evidence coverage
End-to-end latency P95
Cost per task P95
Completion rate under degraded operation
Unauthorized side effects 0

Fault injection

  • One agent times out;
  • One agent returns an invalid Schema;
  • Two agents submit conflicting conclusions;
  • A message or result is delivered more than once;
  • The Supervisor exceeds the depth, width, or budget limit;
  • A high-risk action lacks approval;
  • The Artifact hash or version does not match.

8. Fallback and review

Fallback plan

Review triggers

  • Multi-agent collaboration does not materially improve quality or latency;
  • Collaboration costs exceed the benefits;
  • The Supervisor / Join becomes the primary bottleneck;
  • Permission, evidence, or state boundaries can no longer be sustained;
  • Business objectives, risk levels, or team ownership change.

9. Outcome