Deploy parallel agent teams for multi-lane execution using Claude Code native teams
You are invoking Blitzkrieg Teams: parallel multi-agent dispatch using Claude Code's native team infrastructure (TeamCreate, Task, SendMessage, TaskUpdate, TaskList).
Version: 1.0.0 Last Updated: 2026-02-09 Authority: Oracle 13
Trigger this skill when:
Do NOT use when:
blitz-researchPurpose: Parallel information gathering before a decision or implementation.
Composition: 2-3 agents, all Explore subagent type.
Use when: Multiple questions need answers from different sources simultaneously.
| Slot | Subagent Type | Typical Assignment |
|---|---|---|
| Scout-1 | Explore | Corpus search (CANON, ENGINE, SIGMA docs) |
| Scout-2 | Explore | Configuration audit (dotfiles, launchd, scripts) |
| Scout-3 | Explore | External research (web, API docs, changelogs) |
Example: "Survey the state of all launchd agents, all MCP server configs, and all skill files simultaneously."
blitz-implementPurpose: Parallel implementation across independent file scopes.
Composition: 2-4 agents, all general-purpose subagent type.
Use when: Multiple files/subsystems need changes that do not depend on each other.
| Slot | Subagent Type | Typical Assignment |
|---|---|---|
| Strike-1 | general-purpose | Subsystem A implementation |
| Strike-2 | general-purpose | Subsystem B implementation |
| Strike-3 | general-purpose | Subsystem C implementation |
| Strike-4 | general-purpose | Test/validation harness |
Example: "Update cockpit.sh pane config, write new launchd plist, and create the dispatch script simultaneously."
blitz-mixedPurpose: Research phase followed by implementation phase, or research running alongside implementation.
Composition: 1-2 Explore + 1-2 general-purpose agents.
Use when: Some lanes need discovery while others can begin immediately.
| Slot | Subagent Type | Typical Assignment |
|---|---|---|
| Researcher-1 | Explore | Investigate unknowns, gather requirements |
| Researcher-2 | Explore | Audit current state, find conflicts |
| Implementer-1 | general-purpose | Begin work on known-good lanes |
| Implementer-2 | general-purpose | Pick up implementation after research completes |
Example: "Research the Qdrant MCP configuration while simultaneously implementing the Linear MCP setup."
Before spawning any agents, decompose the work:
Output format (write this to your working memory before proceeding):
BLITZKRIEG DECOMPOSITION
Lane A: <description> | Files: <scope> | Success: <criteria>
Lane B: <description> | Files: <scope> | Success: <criteria>
Lane C: <description> | Files: <scope> | Success: <criteria>
Sync point: <where lanes must converge, if any>
Choose the composition pattern:
blitz-research)blitz-implement)blitz-mixed)Step 1: Create the team
TeamCreate with team_name (e.g., "blitz-implement")
Step 2: Create tasks for each lane
TaskCreate for Lane A with description and success criteria
TaskCreate for Lane B with description and success criteria
(repeat for each lane)
Step 3: Spawn teammates
For each lane, use the Task tool:
- team_name: the team created in Step 1
- subagent_type: "Explore" or "general-purpose" per the composition pattern
- description: the lane's full context including:
a. What to do (the task)
b. Which files are in scope (explicit paths)
c. Which files are OUT of scope (explicit exclusion)
d. Success criteria
e. "Report results via SendMessage to team lead when complete"
Step 4: Assign tasks
TaskUpdate to assign each task to the spawned teammate
While teammates execute:
If a lane is blocked:
When all lanes complete:
git status, relevant tests, grep checksfeat(blitzkrieg): <what was accomplished across lanes>
After integration:
DYN-EXECUTION_STAGING.md (per Execute skill protocol)append_ledger.sh if the directive warrants it| Name | Pattern | Typical Size |
|---|---|---|
blitz-research | Scout Team | 2-3 Explore agents |
blitz-implement | Strike Team | 2-4 code agents |
blitz-mixed | Mixed Team | 1-2 Explore + 1-2 general-purpose |
Use these names consistently. They appear in TaskList output and execution logs.
Blitzkrieg Teams uses Claude Code's native team infrastructure for Claude-to-Claude parallelism. For dispatching to non-Claude constellation agents (Psyche, Ajna, Adjudicator via Codex, Cartographer via Gemini), use the repo's dispatch system:
# Dispatch to non-Claude agents via agent inbox folders
bash orchestration/scripts/dispatch.sh <agent> "<TOPIC>" "<DESC>" "" "<KIND>" "commander"
Hybrid pattern -- Claude team + constellation dispatch:
blitz-implement team for Claude Code lanesdispatch.shagents/commander/inbox/pending/ for RESULT files| Target | Mechanism | Monitor Via |
|---|---|---|
| Claude Code agents | TeamCreate + Task + SendMessage | TaskList |
| Psyche (GPT-5.3-codex) | dispatch.sh -> agents/psyche/inbox/ | RESULT files in commander inbox |
| Ajna (Kimi K2.5) | dispatch.sh -> agents/ajna/inbox/ | RESULT files in commander inbox |
| Adjudicator (Codex CLI) | dispatch.sh -> agents/adjudicator/inbox/ | RESULT files in commander inbox |
| Cartographer (Gemini CLI) | dispatch.sh -> agents/cartographer/inbox/ | RESULT files in commander inbox |
Blitzkrieg for sequential work: If Lane B needs Lane A's output, they are not parallel. Use the Execute skill with ordered steps instead.
Overlapping file scopes: Two agents writing to the same file guarantees conflicts. Restructure the decomposition until scopes are disjoint.
Fire-and-forget spawning: Spawning agents without monitoring is resource waste. Always track via TaskList and collect via SendMessage.
Git operations by teammates: Only the team lead commits. Teammates that run git commands will corrupt the index. This rail is absolute.
Spawning without decomposition: Jumping to TeamCreate before writing the lane decomposition produces vague scopes and redundant work.
Over-parallelism: 4 agents doing 15-minute tasks each is not faster than 1 agent doing 4 tasks sequentially if the coordination overhead exceeds the time saved. Minimum useful parallelism: each lane takes 5+ minutes independently.
Ignoring partial failure: If one lane fails, do not discard the others. Commit successful lanes, re-plan the failed one.
1. TeamCreate(team_name="blitz-implement")
2. TaskCreate(description="Lane A: ...", team_name="blitz-implement")
3. TaskCreate(description="Lane B: ...", team_name="blitz-implement")
4. Task(team_name="blitz-implement", subagent_type="general-purpose", description="Lane A context...")
5. Task(team_name="blitz-implement", subagent_type="general-purpose", description="Lane B context...")
6. TaskUpdate(task_id=<lane_a_id>, assignee=<agent_a_name>)
7. TaskUpdate(task_id=<lane_b_id>, assignee=<agent_b_name>)
8. [Monitor] TaskList(team_name="blitz-implement")
9. [Collect] Read SendMessage results from teammates
10. [Verify] Check artifacts, run tests
11. [Commit] git add + git commit (team lead only)
12. [Shutdown] SendMessage(type="shutdown_request") to each teammate
.claude/skills/plan.md -- Plan generation (upstream; BLITZKRIEG mode feeds into this skill).claude/skills/execute.md -- Execution tracking (downstream; logs and completion protocol).claude/skills/triage.md -- Inbox triage (monitoring RESULT files from constellation dispatch)orchestration/scripts/dispatch.sh -- Constellation dispatch to non-Claude agentsorchestration/state/REF-NEO_BLITZKRIEG_BUILDOUT.md -- Blitzkrieg architectural referenceorchestration/state/DYN-EXECUTION_STAGING.md -- Execution tracking ledgerCLAUDE.md -- Constitutional rules and Directive Completion ProtocolBlitzkrieg Teams Skill v1.0.0 | Syncrescendence