compound-eng-orchestrating-swarms（复合工程编排集群）

Swarm orchestration

Primitives

Agents, teams, teammates, leaders, tasks, inboxes, messages, backends — see primitives.md for definitions and the file-system layout.

Two Ways to Spawn Agents

Subagent (short-lived, returns result):

Task({ subagent_type: "Explore", description: "Find auth files", prompt: "..." })

Teammate (persistent, communicates via inbox):

Teammate({ operation: "spawnTeam", team_name: "my-project" })
Task({ team_name: "my-project", name: "worker", subagent_type: "general-purpose",
       prompt: "...", run_in_background: true })

For detailed agent type descriptions, see agent-types.md.

Parallel Fan-Out (for independent work)

When dispatching multiple read-only or worktree-isolated agents whose work is independent, issue all Task calls in a SINGLE assistant message. Sequential dispatch across separate messages serializes what should run concurrently. Opus 4.7 does not parallelize by default -- state it explicitly.

// Correct: one message, multiple Task tool uses
Task({ subagent_type: "security-sentinel", ... })
Task({ subagent_type: "performance-oracle", ... })
Task({ subagent_type: "architecture-strategist", ... })

Sequential dispatch (each Task in its own message, waiting on the previous to return) is a serialization bug, not a coordination pattern. If agents truly depend on each other's output, that is a pipeline -- see Coordination Models below.

Bounded parallelism when the harness caps active subagents. Single-message fan-out (above) tells Opus to dispatch in parallel; the harness then decides how many to *run* concurrently. When the harness accepts the dispatch but caps active execution, queue the overflow rather than failing. Dispatch as many as the harness accepts in the first batch, treat transient capacity-related spawn errors as backpressure (any retryable error indicating the limiter rejected the dispatch — exact wording varies across harness versions and platforms; do not pattern-match on a fixed string list), and re-dispatch queued agents as active ones complete. Record an agent as failed only after a successful dispatch times out or returns an error, or when dispatch fails for a non-capacity reason (bad tool name, malformed prompt, missing permission). The fan-out is still parallel — it is just rate-capped to whatever the harness can run concurrently.

Quick Reference

For copy-paste spawn/message/task/shutdown snippets, load quick-reference.md.

Dispatch Discipline

Rules for when and how to dispatch agents. Getting these wrong wastes tokens and creates hard-to-debug failures.

When to dispatch a team vs. do it yourself:

Assess 5 signals: file count, module span, dependency chain, risk surface, parallelism potential. If 3+ fall in the "complex" column, dispatch a team. Below 3, do it yourself. When in doubt, prefer the simple path -- team overhead is only justified when parallelism provides a real speedup.

Task description template (for every dispatched task):

Every task prompt must include these fields to prevent integration failures:

• Objective: what to accomplish (one sentence)
• Owned Files: files this agent creates or modifies (exclusive -- no file assigned to multiple agents)
• Interface Contracts: what to import from other agents' work, what to export for downstream agents
• Acceptance Criteria: how the agent knows the task is correct
• Out of Scope: what NOT to touch, even if it looks related

Cardinal rule: one owner per file. When files must be shared, designate a single owner; other agents send change requests, owner applies sequentially. If an upstream dependency isn't ready yet, write a stub/mock so downstream work can continue unblocked.

No parallel implementation agents (without worktrees):

Implementation agents share state via git by default, so parallel dispatch causes overwrites. Use isolation: "worktree" to give each agent its own copy. Without worktrees, dispatch implementation agents sequentially. Review, research, and analysis agents are always safe to parallelize (read-only).

Pre-dispatch file-intersection check -- operationalize the one-owner-per-file rule with a runnable safety gate before every parallel dispatch:

1. Collect each unit's declared Owned Files / Test Paths / Modify Paths from its task spec.
2. Build a {file → unit} map. If any file appears under more than one unit, the dispatch is unsafe. Quick check on Markdown task specs:

   grep -h "^Owned Files:" -A 20 tasks/*.md | grep -v "^Owned Files:" | grep -v "^--$" | sort | uniq -d

Any output is an overlapping file path that needs resolution.

1. On overlap: either downgrade to serial (log the overlap and the reason), or assign worktree isolation (isolation: "worktree" per agent), or rewrite unit boundaries so files become exclusive.
2. Even with no declared overlap, include this constraint verbatim in every parallel-dispatch prompt: *"Do not run git add, git commit, or the project's test suite while other parallel agents are active -- you'd race on the git index or thrash the test cache. Stage changes for the orchestrator to commit after integration."*

The intersection check catches silent conflicts the controller misses at plan time; the dispatch-prompt constraint catches them when a unit's file list was incomplete.

Preset team compositions: Start from a named preset before designing a custom team. See team-compositions.md for the full table (Review / Debug / Feature / Fullstack / Migration / Security / Research), the cardinal subagent_type rule (read-only agents cannot implement), and custom-team guidelines. Use the smallest preset that covers all required dimensions — overlap between reviewers is a sizing signal to redefine focus areas, not add more agents.

Model selection by task complexity:

Handoff protocol -- structured agent-to-agent transfers:

When passing work between agents (leader→implementer, implementer→reviewer, reviewer→leader), include:

1. Context: what was done, relevant files, constraints discovered
2. Deliverable: specific output expected from the receiving agent
3. Acceptance criteria: how the receiving agent knows the work is correct

The controller reads all tasks from the plan upfront and provides full task text directly to subagents. Never make subagents read plan files themselves -- they waste tokens navigating, may read different versions, and inherit unclear context. Paste the task content into the prompt. See handoff-templates.md for QA FAIL and Escalation Report formats.

Standardize implementer status signals:

Include the four statuses defined in ia-verification-before-completion (DONE, DONE_WITH_CONCERNS, BLOCKED, NEEDS_CONTEXT) in every teammate prompt so they know the reporting format. Expect teammates to report one. BLOCKED responses get further triage via the decision tree below.

BLOCKED triage decision tree -- when a teammate reports BLOCKED, classify the root cause before acting. Never retry the same prompt on the same model without changing a variable.

Never ignore an escalation. Never force the same agent to retry without changing at least one variable (context, model, or task scope).

Two-stage review gate on subagent outputs:

Verify spec compliance first: does the output match what was requested? Only then evaluate quality. A beautifully written solution to the wrong problem is still wrong. Structure review as two explicit passes -- pass 1 rejects on spec mismatch without reading further, pass 2 assesses correctness and quality on spec-compliant outputs.

QA retry loop:

Max 3 attempts per task. After each QA failure, pass structured feedback to the implementer using the QA FAIL template. After 3 failures, mark the task as blocked, continue the pipeline (don't halt everything), and let final integration catch remaining issues. Counter resets when advancing to the next task.

Integration Rules

Post-integration verification -- after all agents return: check overlapping file edits, review for conflicting approaches, run full test suite.

Spawned-session behavior -- when a skill runs inside an orchestrated pipeline (as a subagent, not user-invoked), suppress interactive prompts: do not use AskUserQuestion, auto-choose the conservative/safe default, skip upgrade checks and telemetry. Focus on completing the task and reporting results via prose output. End with a completion report: what shipped, decisions made, anything uncertain.

Context Carry-Forward

After each turn, five strategies exist for moving context forward: Continue, Rewind, /compact, Subagent, /clear+brief. Choose deliberately — the default "Continue" is rarely best, and Rewind is strictly better than "correcting in place" after a failed attempt. See context-carry-forward.md for the full decision table and rationale.

Coordination Models

Two approaches to multi-agent coordination exist. Choose based on the work pattern:

For most work, start with stateless handoffs. Graduate to stateful coordination only when parallelism provides a real speedup and you have worktree isolation to prevent file conflicts.

Dispatch Anti-Patterns

Before designing any multi-agent workflow, check it against the four named failure modes in dispatch-anti-patterns.md: router persona, persona calls persona, sequential paraphraser, deep persona trees. Rule of thumb: if the proposed swarm has more coordinator roles than worker roles, collapse it.

Anti-Sycophancy and Resilience

When dispatching judge panels, running parallel reviewers, or iterating on subjective evaluations, load anti-sycophancy.md — cold-start isolation, fresh instances per round, label randomization, convergence detection.

When designing multi-agent workflows that must survive partial failure, load resilience-patterns.md — cascade prevention (timeouts, circuit breakers, bulkheads), failure classification (retry vs reassign vs escalate), mid-pipeline compensation for irreversible side effects, post-failure synthesis of partial results.

Verify

• All tasks in terminal state (completed or blocked)
• No orphaned teammates (git worktree list shows no stale entries)
• Overlapping file edits reviewed and merged
• Full test suite passes post-integration

References