# Agent Teams Command

> Use when work has genuinely independent streams or distinct builder, evaluator, domain, and integration roles that require bounded multi-agent command.

- **Type:** Skill
- **Install:** `agentstack add skill-mark393295827-third-brain-v7-skills-agent-teams-command`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [Mark393295827](https://agentstack.voostack.com/s/mark393295827)
- **Installs:** 0
- **Category:** [Productivity](https://agentstack.voostack.com/c/productivity)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [Mark393295827](https://github.com/Mark393295827)
- **Source:** https://github.com/Mark393295827/third-brain-v7-skills/tree/main/skills/agent-teams-command
- **Website:** https://github.com/Mark393295827/third-brain-v5-skills/tree/master

## Install

```sh
agentstack add skill-mark393295827-third-brain-v7-skills-agent-teams-command
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Agent Teams Command

  A mission with independently ownable workstreams, interfaces, permissions, budgets, and an integration owner.
  An isolated worker program, typed IPC ledger, serial integration, cleanup, and evidence receipts.
  Integrated artifacts pass mission checks and ownership, review, rollback, and cleanup gates.
  Graph-schema design, parallelism without net value, overlapping writers, or worker self-certification.

Strategic intent and final integration remain serial; owned execution may run
in parallel. `graph-engineering` may define dependency topology, but this skill
owns worker processes, exclusive territories, typed IPC, integration, and
cleanup. A Graph node is not automatically a teammate. Use the Ender lens for
commander understanding, Palantir for operational objects/actions, and von
Neumann for executable, inspectable process architecture. Detailed patterns:
`references/ender-palantir-command-patterns.md`; examples:
`references/classic-campaigns.md`.

## Usage Template

Provide: mission, non-goals, workstreams, dependencies, files/systems, acceptance criteria, permissions, review bandwidth, runtime capabilities, budget, and integration owner.

## Workflow

Admit a team only when at least two workstreams can proceed with low coordination, or distinct roles materially improve evaluation/safety. Calculate orchestration tax: setup, context duplication, IPC, merge conflict, review, and cleanup. If one process can finish within the same review budget, keep one process.

If dependencies and joins are the only complexity, use `graph-engineering`
without recruiting a team. Add workers only for nodes that need isolated
context, distinct ownership, or independent judgment.

Resolve commander intent, object/state vocabulary, ownership, dependency direction, verifier, join gate, and permission boundary before recruiting. Return `NEEDS_INPUT` when an irreversible business decision or authority cannot be discovered locally. Workers may probe implementation unknowns only inside their territory.

1. Write a command program: objective, non-goals, finite actions, commander, checkpoint cadence, interrupt policy, state/artifact paths, IPC schema, allowed/denied tools, verifier, stop, recovery, and promotion boundary.
2. Atomically decompose work by interface/territory. Each task has one owner, inputs, output artifact, dependencies, definition of done, verifier, budget, and blast radius.
3. Select the smallest topology: maker-checker, manager-workers, or specialist pipeline. If a validated Graph exists, map only `agent` or `agent-team` nodes to workers; deterministic, Loop, subgraph, and human-gate nodes keep their own contracts. Route workers by required capabilities and runtime policy, never fixed model names.
4. Isolate mutable work with separate worktrees/branches or disjoint files. Shared schemas/contracts are commander-owned until published.
5. Use typed IPC: `{task_id, state, artifact, evidence, decision, unknowns, dependency, next_action}`. Messages change state; status chatter does not.
6. Require workers to run `state + evidence -> next action -> verifier -> next state | stop | escalate` and checkpoint after each material action.
7. Monitor attention and review budgets, repeated failures, idle dependencies, ownership violations, and architecture drift. Rebalance or serialize when coordination cost rises.
8. Integrate one verified workstream at a time in dependency order. The integration owner runs broader regression/eval checks and records accepted/rejected artifacts.
9. Run an independent reviewer/red team for consequential behavior. Human approval precedes production, publication, spending, destructive/shared-state actions, credentials, and policy changes; prepare rollback first.
10. Close workers, remove temporary worktrees/state, reconcile tasks, preserve receipts, and extract only promotion-gated reusable patterns.

Hooks are optional executable infrastructure, not prose. Configure one only when its referenced program exists, has tests, and emits a verified receipt; never copy illustrative hook commands as if installed.

Evaluate mission outcome, per-workstream evidence, Graph edge/join compatibility
when used, ownership compliance, integration diff, regression suite,
reviewability, residual risk, and cleanup. Quiet green checks are insufficient
if the commander cannot explain architecture delta and rollback.

`max_attempts: 2` per failed task/integration signature. Retry the failed unit only after changing diagnosis, owner, scope, or strategy. Stop/rebalance on repeated signature, ownership conflict, rising orchestration tax, exhausted review budget, or `NO_PROGRESS`.

Persist `{run_id, status, attempt, budget, evidence, unknowns, last_error, next_action}` plus mission/contract version, command board, task graph, ownership map, worker/worktree registry, IPC events, attention budget, approvals, integration ledger, rollback points, and cleanup receipt.

## Failure Protocol

- `NEEDS_INPUT`: intent, ownership, verifier, join, or authority is unresolved; do not launch.
- `BLOCKED_DEPENDENCY`: checkpoint the worker and advance only independent tasks.
- `BLOCKED_PERMISSION`: stop the affected stream and request approval.
- `VERIFY_FAILED`: reject the artifact at its integration gate; preserve evidence.
- `NO_PROGRESS`: the same signature repeats after changed strategy. `max_attempts: 2`.
- `BUDGET_STOP`: interrupt workers, checkpoint, and emit a partial integration/cleanup plan.

## Output Contract

Return `status`, `result` (mission and integrated artifacts), `evidence` (task/integration/cleanup receipts), `unknowns`, and `next_action` including approval or rollback.

## Edge Cases

- Two workers need the same schema file: commander publishes the contract first; serialize schema edits and let workers consume it read-only.
- A Graph has six nodes but only two require agent judgment: recruit two
  workers, not six; deterministic nodes remain scheduler-owned.
- All workers report success but integration fails: mission status is `VERIFY_FAILED`; reject incompatible artifacts rather than averaging reports.

## Success Metrics

- Parallel work reduces elapsed time or improves independent evaluation beyond its coordination cost.
- Ownership is exclusive, IPC is typed, and integration is serial and evidence-gated.
- Final state is reproducible, reviewable, recoverable, and clean.

## Quality Gates

- [ ] Team admission and orchestration-tax calculation justify multi-agent use.
- [ ] Every task has one owner, artifact, verifier, budget, and stop condition.
- [ ] Independent review, approval, and rollback match risk.
- [ ] Integration and cleanup receipts match actual repository/runtime state.
- [ ] Capability routing contains no durable model binding.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Mark393295827](https://github.com/Mark393295827)
- **Source:** [Mark393295827/third-brain-v7-skills](https://github.com/Mark393295827/third-brain-v7-skills)
- **License:** MIT
- **Homepage:** https://github.com/Mark393295827/third-brain-v5-skills/tree/master

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-mark393295827-third-brain-v7-skills-agent-teams-command
- Seller: https://agentstack.voostack.com/s/mark393295827
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.