Pattern matrix/White Paper/Collaboration

Collaboration pays off through task scale, specialist boundaries, independent review, or fault isolation. Adding participants also adds hand-off loss, resource conflicts, and goal drift. This module defines those boundaries and shows how the six collaboration patterns fit together.

Three collaboration planes

The third plane often remains in meetings, chat, and individual memory. Decisions, evidence, and boundaries that affect later work belong in versioned assets; raw conversation does not need to be copied wholesale.

Design semantics and runtime primitives

Complex collaboration graphs often reduce to serial, parallel, and routing edges. These primitives describe connectivity. Loops, hierarchy, and orchestration retain higher-level semantics about global state, review, reassignment, and acceptance.

Hierarchical Delegation may compile to routing, parallel calls, and gather. Adversarial Review may compile to generate, review, adjudicate, and conditional loop. Similar edges still carry different responsibility. ADPS calls this topology lowering. Runtime records should preserve the design pattern, roles, contracts, and acceptance references.

Four layers of collaboration design

An order flow may select payment, inventory, and notification agents at runtime. It remains dynamic orchestration while one interpreter owns the plan and gathers results. C6 begins when payment publishes an event, inventory and notification subscribe under local rules, and no node knows the complete flow. Ownership of the plan, not predefinition, is the criterion.

Topology governance matrix

Topology describes how control unfolds; it does not replace authorization or audit. Effective authority narrows across the user, current task, agent role, tool, and resource.

Handoff Contract

Conversation history does not reliably express binding decisions, rejected paths, authority, or acceptance criteria.

handoff_id: h_01K...
goal: preserve the compatibility interface
from_role: requirements-agent
to_role: implementation-agent
artifacts:
  - uri: artifact://spec/417
    version: sha256:...
decisions:
  - choice: keep the public schema
    evidence: adr://23
authority:
  allowed_tools: [repo_read, patch_write]
  resource_scope: repo://service-a
acceptance:
  checks: [unit_tests, contract_tests]

A Handoff Contract transfers goal, artifacts, decisions, responsibility, authority, and acceptance. The receiver explicitly accepts or rejects it, and temporary authority from the previous leg is reclaimed.

Same-stack, heterogeneous, and cross-session collaboration

Agents in one session share a runtime but still need constrained context and tools. Separate sessions or worktrees isolate files, not requirement IDs, shared configuration, databases, or external environments. The scheduler must declare a write-conflict domain before parallel execution.

Heterogeneous agents may use an upper-level coordinator or distributed capability discovery and message forwarding. Both designs require capability descriptions, identity, status, timeout, idempotency, and traceable hand-off.

Independent review and feedback

Separating generator and reviewer reduces self-review bias but can sever execution feedback. Review artifacts should carry evidence, risks, conditions, and revalidation requirements. C3 sets conditions, C4 transfers them, and X1/X2 verify the external result.

Hook composition

Hooks can advance stages, enforce policy decisions, record observation events, or preserve state for recovery. G5 keeps its historical identifier for unavoidable governance enforcement points. Cross-module use is documented in Hook Composition.

From exploration to production pinning

Development can try new task splits and topologies. Test and staging progressively pin agents, models, tools, policies, and the main runtime graph. Production keeps dynamic choices only inside evaluated boundaries. Version changes require new evidence.

When multiple agents are justified

1. A long task inflates one context until early detail interferes with current judgement.
2. Subtasks require different skills, tools, data, or authority.
3. A high-risk artifact needs structurally independent review.
4. Work can be partitioned safely and wall-clock gains exceed communication and aggregation cost.

If one agent with clear skills can complete the task, the single-agent baseline is usually steadier.

Verification

Workshop record

This overview incorporates the first Collaboration Module Workshop on 25 August 2026. Hosts: Haili Zhang and Jia Huang. Core workshop guests: Dong Zhang and Wei Wang.

Read the workshop record · Human-Agent collaboration · Abstraction and Reconstruction