The philosophy and practical benefits of agent fungibility in multi-agent software development. Why homogeneous, interchangeable agents outperform specialized role-based systems at scale.
日本語の概要は準備中です。原文の説明を表示しています。
Orchestrates NTM tmux agent swarms and robot APIs. Use when spawning/sending panes, robot state, triaging work, locks/mail, safety, pipelines, serve, code-first/batch-verify waves, or NTM errors.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
NTM is a tmux control plane for multi-agent work. This entrypoint is the operating kernel. Load one linked reference only when the current action needs it.
Discover the live contract, choose the smallest non-interactive surface, act, then verify on the read surface that owns the resulting state.
Never infer success from command acceptance alone.
AGENTS.md and README.md.ntm --robot-capabilities --capability-compact
ntm --robot-capabilities --capability-command=send --capability-compact
ntm --robot-capabilities --capability-search=interrupt --capability-compact
ntm --robot-snapshot
For older NTM builds without capability filters, use
ntm --robot-capabilities and narrow locally.
DISCOVER -> SNAPSHOT -> SELECT -> SCOPE -> EXECUTE -> VERIFY -> CLEAN UP
Re-snapshot after a cursor expiry, degraded-source recovery, or material state transition.
Before a state change, answer:
- Target: <project/session and canonical pane refs>
- Contract: <capability/schema checked>
- Evidence before: <cursor, source health, state>
- Ownership: <bead assignee and file reservations>
- Safety: <user pane inclusion, approvals, blast radius>
- Postcondition: <owning read surface>
- Recovery: <retry, interrupt, smart restart, restore, handoff>
If any field is unknown, stay read-only.
| Intent | First surface | Load |
|---|---|---|
| Install, configure, environment | ntm deps, config | CONFIG, ENV-VARS |
| Spawn/add/adopt agents | ntm spawn, ntm add | SPAWN |
| Send or batch prompts | --robot-send | SEND |
| Inspect state/output | snapshot, tail, inspect | ROBOT-MODE |
| Pick and assign work | work triage, assign | WORK-AND-ASSIGN |
| Coordinate files/agents | locks, mail, coordinator | INTEGRATIONS |
| Recover or resume | diagnose, checkpoint, handoff | TROUBLESHOOTING, DURABILITY |
| Run repeatable phases | pipeline | PIPELINES |
| Run reasoning ensembles | ensemble | ENSEMBLE |
| Operate HTTP/WebSocket API | serve | SERVE |
| Review policy/approvals | safety, policy, approve | SAFETY |
| Use human dashboard | dashboard/palette | DASHBOARD |
| Find exact CLI syntax | live capabilities first | COMMANDS |
| Run high-throughput coding waves | beads policy + orchestrator | CODE-FIRST-BATCH-VERIFY |
For an already-running multi-pane swarm, use /vibing-with-ntm for the tending
loop and return here for exact command contracts.
When many panes share one repo and one expensive build path, run the two-phase pump: parallel code-first waves (real code + real tests, syntax gate max, commit immediately) followed by one central batch-verify pass that alone closes work with revision-bound evidence.
PHASE 1 (all panes, parallel): claim -> code+tests -> syntax gate -> commit -> batch_pending -> next
| trigger: ready-pool dry | debt ceiling | articulation point | scope frontier | time/risk | rate dip
v
PHASE 2 (orchestrator, once): commit-flush -> ONE verify over git-derived scope
-> compile errors first -> cluster failures by file -> rework (same assignee)
-> re-run green -> gate + close with evidence -> dependents unblock -> next wave
Full mechanics, tracker policy shape, triggers, and enforcement: CODE-FIRST-BATCH-VERIFY.
The pump is only safe on top of honest credit: process artifacts gate named
features or don't exist; refusal-only work never closes a positive-capability
item; closures cite evidence; metrics predeclare denominators. That incentive
layer — including how to encode it into the beads themselves so every swarm
agent sees it at claim time — is the vibing-with-ntm skill's HONEST-CREDIT
reference; its DOCTRINE-BOOTSTRAP reference is the step-by-step setup
(AGENTS.md law, tracker policy, root doctrine bead, AC templates,
enforcement canary). Encode both doctrines in the target repo's AGENTS.md
before the first wave.
Use canonical selectors returned by live robot state:
%N: stable tmux pane ID.W.P: explicit window and pane.N: pane index in one-window sessions; window index in multi-window
sessions.Do not assume the user pane is index 0. Both dispatch surfaces exclude
user-typed panes by default, but their --all flags differ: robot --all
opts the user pane in, while shell ntm send --all broadens to all agent
panes and reaches the user pane only with --include-user. Prefer %N or
W.P for any state-changing single-pane action.
A malformed, missing, or ambiguous selector is a command error, not an empty successful target set.
A session is a project name or project--label. The project directory must
resolve to the repository NTM, Beads, and Agent Mail should share. Check
projects_base, session labels, and the resolved absolute path before
claiming work or reserving files.
Do not repair resolution by guessing symlinks or moving repositories. Inspect configuration and follow repo policy.
ntm view, dashboard, palette, and other TUIs in automation.--robot-* output; use --robot-format=toon and
--robot-verbosity=terse when supported and context is tight.success as authoritative, then inspect error_code and hint.0 means success, 1 means command error, and 2 means
unavailable/NOT_IMPLEMENTED.[], never inferred from omission.Treat the live capability catalog as the robot-surface count; do not copy a number into automation guidance:
ntm --robot-capabilities | jq '.commands | length'
ntm respawn <session> and ntm kill <session> ask for confirmation. An
orchestrator must use --force only after its target and blast radius are
explicitly verified; otherwise the command will wait at a prompt and silently
stall automation.
For model-qualified launches, use N:model:effort where the agent supports an
effort hint, for example --cod=2:gpt-5.6-terra:high. Before using either
qualifier in a custom [agents] command, ensure the Go template contains
{{.Model}} (or {{.ModelAlias}}) and, for effort, {{.ReasoningEffort}}.
NTM rejects a requested qualifier that its template would ignore; fix the
template rather than assuming the launch received the setting.
A detached tmux session may begin at 80x24; a wide swarm can leave each pane too small for agent dialogs and even make dialog detection miss wrapped options. Before spawning or tending a large detached swarm, size the window:
tmux resize-window -t <session> -x 420 -y 110
Re-check the actual pane geometry after layout changes; this is an operator environment constraint, not an agent failure.
| Mutation | Required read-back |
|---|---|
| Spawn/add/restart | session status plus pane PID/current command |
| Send/interrupt | target receipt plus fresh pane capture |
| Bead claim/assign/close | native br show/ready plus assignment state |
| File reservation | Agent Mail reservation result/conflict list |
| Pipeline run/resume/cancel | pipeline state/run ID |
| Lock acquire/release | lock owner/lease state |
| Checkpoint/handoff | artifact metadata and restore target |
| Safety approval | policy/token state |
| Event-producing action | events/digest/attention cursor movement |
Use git, Beads, process state, or emitted artifacts as independent evidence when the action is meant to change them.
INVALID_FLAG appears; its
hint field now carries a did-you-mean suggestion for near-miss flags.An NTM action is complete only when:
Grep the references instead of loading them whole:
REFS=.claude/skills/ntm/references
# Find the exact flag/command contract
grep -niE "<flag-or-subcommand>" "$REFS"/COMMANDS.md "$REFS"/ROBOT-MODE.md
# Find pump mechanics by keyword
grep -niE "trigger|commit-flush|rework|gate|build-kill|revision" \
"$REFS"/CODE-FIRST-BATCH-VERIFY.md
# Find closure/claim discipline
grep -niE "actor|close|ready|claim|dep add|cycles" "$REFS"/WORK-AND-ASSIGN.md
# Find recovery for an error string
grep -niE "<error-text>" "$REFS"/TROUBLESHOOTING.md "$REFS"/DURABILITY.md
/vibing-with-ntm: tend active swarms/beads-br, /beads-bv: tracker mechanics and graph triage/agent-mail: coordination primitives/brennerbot-with-ntm: hypothesis research sessions/open-beads-weighted-tmux-agent-sessions: backlog-weighted spawningまだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
The philosophy and practical benefits of agent fungibility in multi-agent software development. Why homogeneous, interchangeable agents outperform specialized role-based systems at scale.
日本語の概要は準備中です。原文の説明を表示しています。
MCP Agent Mail for multi-agent coordination. Use when agents need file locks, messaging, inboxes, or conflict prevention. Handles macro_start_session, file_reservation_paths, send_message, threading, pre-commit guards.
日本語の概要は準備中です。原文の説明を表示しています。
Jeffrey Emanuel's multi-agent implementation workflow using NTM, Agent Mail, Beads, and BV. The execution phase that follows planning and bead creation. Includes exact prompts used.
日本語の概要は準備中です。原文の説明を表示しています。
Convert markdown plans into beads with dependencies using br CLI. Use when creating task graphs, polishing beads before implementation, or bridging planning to agent swarm execution.
日本語の概要は準備中です。原文の説明を表示しています。
Beads Viewer - Graph-aware triage engine for Beads projects. Computes PageRank, betweenness, critical path, and cycles. Use --robot-* flags for AI agents.
日本語の概要は準備中です。原文の説明を表示しています。
Manage AI coding CLI accounts with sub-100ms switching. Use when hitting rate limits on Claude Max, GPT Pro, or Gemini Ultra subscriptions and need instant account swapping without browser OAuth.
日本語の概要は準備中です。原文の説明を表示しています。