GitHub - battysh/batty: Supervised agent execution for software teams. Kanban-driven, tmux-native, test-gated.

7 min read Original article ↗

Batty

Self-improving hierarchical agent teams for software development.

Define a team in YAML, launch it in tmux, and let Batty handle the happy path: dispatch work, isolate engineers in worktrees, verify completions, and auto-merge safe changes back to `main`.

CI crates.io Downloads MIT License Built with Rust

Quick Start · Docs · GitHub

Batty is a control plane for agent software teams. Instead of one overloaded coding agent, you define roles such as architect, manager, and engineers; Batty launches them through typed SDK protocols or shim-backed PTYs, routes work between roles, tracks the board, keeps engineer work isolated in git worktrees, and closes the loop with verification and auto-merge.

How Batty works: Define → Supervise → Execute → Verify → Deliver
How Batty works: Define → Supervise → Execute → Verify → Deliver

Quick Start

cargo install batty-cli
batty init
batty start
batty attach
batty status

cargo install batty-cli installs the batty binary. After batty init, edit .batty/team_config/team.yaml, start the daemon, attach to the live tmux session, and use a second shell to send the architect the first directive:

batty send architect "Build a small API with auth, tests, and CI."

For the step-by-step setup flow, see docs/getting-started.md.

What v0.10.0 Adds

Batty v0.10.0 closes the autonomous development loop. Type $go in Discord, go to sleep, wake up to merged features.

  • Discord control surface — three-channel bot (#commands, #events, #agents) with $go/$stop/$status/$board commands and rich embeds
  • Closed verification loop — daemon auto-tests completions, retries on failure, merges on green. No agent in the merge path.
  • Notification isolation — daemon chatter stays in the orchestrator log, not in agent PTY context. Agents stay focused on code.
  • Supervisory stall detection — architect and manager roles get the same health monitoring as engineers. No more silent 30-minute stalls.
  • Manager inbox signal shaping — 200 raw messages/session batched into prioritized digests. Manager sees what matters.
  • Hashline-style edit validation — content-hash checks prevent stale-file corruption when multiple agents edit concurrently.
  • 3,080+ tests, up from 2,854 in v0.9.0.

Architecture

User (Discord / Telegram / CLI)
        |
        v
Architect (Claude) ──> Roadmap ──> Board Tasks
        |
        v
Manager (Claude) ──> Review + Merge
        |
        v
Engineers (Codex x3) ──> Worktrees ──> Code + Tests
        |
        v
Daemon ──> Verify ──> Auto-merge ──> main
        |
        v
Discord (#events, #agents, #commands)

The daemon is the control plane. Discord is the recommended monitoring and control surface — three channels with rich embeds, commands, and mobile access. tmux is the agent runtime display (what agents see), not the primary human interface. Each agent uses a typed SDK protocol (Claude: stream-json NDJSON, Codex: JSONL event stream, Kiro: ACP JSON-RPC 2.0) or falls back to the shim-owned PTY runtime.

Features

  • Hierarchical supervision: architect-level planning, manager-level dispatch, and bounded engineer execution.
  • Daemon-owned workflow loop: auto-dispatch, review routing, claim TTLs, merge queueing, verification retries, and board reconciliation.
  • Discord + Telegram: three-channel Discord with rich embeds and commands, single-channel Telegram with the same command surface. Monitor from your phone.
  • Multi-provider support: mix Claude, Codex, Kiro, and other supported agent CLIs per role.
  • Per-worktree isolation: each engineer gets a stable git worktree and fresh task branches without stomping on other engineers.
  • Self-healing runtime: crash respawn, stall detection (all roles), delivery retries, context exhaustion handoffs, and auto-restart.
  • Closed verification loop: engineer completions are auto-tested, retried on failure, and merged on green without human review in the path.
  • Observability: batty status, batty metrics, SQLite telemetry, Grafana dashboards, daemon logs, and board health views.
  • OpenClaw integration: supervisor contract, DTOs, and multi-project event streams for external orchestration.
  • Clean-room workflow: optional barrier groups, verification commands, and parity artifacts for re-implementation work.

Configuration

Batty topology and runtime workflow live in .batty/team_config/team.yaml. This is a complete example with the fields most teams touch in v0.10.0:

name: my-project
agent: claude
workflow_mode: hybrid
use_shim: true
use_sdk_mode: true
auto_respawn_on_crash: true
orchestrator_pane: true
orchestrator_position: left
external_senders: [slack-bridge]
shim_health_check_interval_secs: 30
shim_health_timeout_secs: 90
shim_shutdown_timeout_secs: 10
shim_working_state_timeout_secs: 1800
pending_queue_max_age_secs: 600
event_log_max_bytes: 5242880
retro_min_duration_secs: 900

board:
  rotation_threshold: 20
  auto_dispatch: true
  auto_replenish: true
  worktree_stale_rebase_threshold: 5
  state_reconciliation_interval_secs: 30
  dispatch_stabilization_delay_secs: 30
  dispatch_dedup_window_secs: 60
  dispatch_manual_cooldown_secs: 30

standup:
  interval_secs: 300
  output_lines: 40

automation:
  timeout_nudges: true
  standups: true
  failure_pattern_detection: true
  triage_interventions: true
  review_interventions: true
  owned_task_interventions: true
  manager_dispatch_interventions: true
  architect_utilization_interventions: true
  intervention_idle_grace_secs: 60
  intervention_cooldown_secs: 300
  utilization_recovery_interval_secs: 900
  commit_before_reset: true

workflow_policy:
  wip_limit_per_engineer: 1
  review_nudge_threshold_secs: 1800
  review_timeout_secs: 7200
  stall_threshold_secs: 120
  max_stall_restarts: 5
  context_pressure_threshold: 100
  context_pressure_threshold_bytes: 512000
  context_pressure_restart_delay_secs: 120
  auto_commit_on_restart: true
  context_handoff_enabled: true
  handoff_screen_history: 20
  verification:
    max_iterations: 5
    auto_run_tests: true
    require_evidence: true
    test_command: cargo test
  claim_ttl:
    default_secs: 1800
    critical_secs: 900
    max_extensions: 2
    progress_check_interval_secs: 120
    warning_secs: 300
  auto_merge:
    enabled: true
    max_diff_lines: 200
    max_files_changed: 5
    max_modules_touched: 2
    confidence_threshold: 0.8
    require_tests_pass: true
    post_merge_verify: true

grafana:
  enabled: true
  port: 3000

roles:
  - name: human
    role_type: user
    channel: telegram
    channel_config:
      provider: openclaw
      target: "123456789"
    talks_to: [architect]

  - name: architect
    role_type: architect
    agent: claude
    prompt: batty_architect.md
    posture: orchestrator
    model_class: frontier
    talks_to: [human, manager]

  - name: manager
    role_type: manager
    agent: claude
    prompt: batty_manager.md
    posture: orchestrator
    model_class: frontier
    talks_to: [architect, engineer]

  - name: engineer
    role_type: engineer
    agent: codex
    instances: 3
    prompt: batty_engineer.md
    posture: deep_worker
    model_class: standard
    use_worktrees: true
    talks_to: [manager]

See docs/config-reference.md for the hand-written team.yaml guide and docs/reference/config.md for the lower-level .batty/config.toml runtime defaults.

Monitoring

These are the day-to-day commands that matter once the team is running:

batty status
batty board health
batty metrics
batty telemetry summary
batty grafana status
  • batty status gives the quickest liveness view.
  • batty board health shows stale tasks, dependency problems, and queue health.
  • batty metrics and batty telemetry summarize throughput, review latency, and agent utilization.
  • batty grafana setup|status|open manages the built-in dashboard.

Troubleshooting

  • Claude or Codex stalls: keep auto_respawn_on_crash: true; inspect .batty/daemon.log, batty status, and batty doctor for restart evidence.
  • Cargo lock contention: use engineer worktrees with shared targets; avoid ad hoc target/ directories inside each worktree.
  • OAuth/auth confusion: prefer current CLI auth flows and avoid relying on stale API-key-only setups.
  • Disk pressure: use batty doctor --fix, archive done tasks, and clean unused worktrees if long-lived teams accumulate state.

More operational guidance lives in docs/troubleshooting.md.

Documentation

Highlights

  • Hierarchical agent teams instead of one overloaded coding agent
  • SDK mode by default for Claude Code, Codex CLI, and Kiro CLI
  • PTY shim fallback when typed protocol support is unavailable
  • tmux-backed visibility with persistent panes and resume support
  • Stable per-engineer worktrees with fresh task branches
  • Auto-dispatch, verification, review routing, and auto-merge
  • SQLite telemetry, Grafana monitoring, and board health reporting

Telegram Integration

Batty can expose a human endpoint over Telegram through a user role. This is useful when you want the team to keep running in tmux while you send direction or receive updates from your phone.

The fastest path is:

batty init --template simple
batty telegram
batty stop && batty start

batty telegram guides you through:

  • creating or reusing a bot token from @BotFather
  • discovering your numeric Telegram user ID
  • sending a verification message
  • updating .batty/team_config/team.yaml with the Telegram channel config

After setup, the user role in team.yaml will look like this:

- name: human
  role_type: user
  channel: telegram
  talks_to: [architect]
  channel_config:
    provider: telegram
    target: "123456789"
    bot_token: "<telegram-bot-token>"
    allowed_user_ids: [123456789]

Notes:

  • You must DM the bot first in Telegram before it can send you messages.
  • bot_token can also come from BATTY_TELEGRAM_BOT_TOKEN instead of being stored in team.yaml.
  • The built-in simple, large, software, and batty templates already include a Telegram-ready user role.

Grafana Monitoring

Batty includes a bundled Grafana dashboard template for long-running team sessions. Use it alongside batty metrics and batty telemetry when you want more than a point-in-time CLI snapshot.

The dashboard JSON lives at src/team/grafana/dashboard.json. Import it into Grafana and point the datasource at .batty/telemetry.db.

Pre-configured alerts:

Alert Detects
Agent Stall Agent silent past threshold
Delivery Failure Spike Message delivery failures climbing
Pipeline Starvation Not enough work in the pipeline
High Failure Rate Tasks failing above threshold
Context Exhaustion Agent context window nearly full
Session Idle Entire team idle too long

Docs and Links

License

MIT