telegram

Agent Bridge

AI 에이전트 팀을 만들어서 회사 업무를 자동화하세요.

Agent Bridge는 Claude Code와 Codex 같은 AI 코딩 에이전트를 "직원"처럼 여러 명 동시에 띄워서, 각자 맡은 업무를 수행하게 하는 로컬 오케스트레이션 플랫폼입니다.

5명이 할 일을 에이전트 20명이 대신합니다. 설정부터 유지보수까지 전부 자연어로.

Current release: v0.17.9 (2026-07-13) — stable line (LTS). A fleet-down split-brain patch on the v0.17 stable line (stability + security patches only, no new features — features land on the separate v0.18.0-beta line), the head that --channel stable (fresh installs) and GitHub Latest resolve to globally — and that --channel lts (production pins) resolves to from a source whose LTS_SERIES is 0.17 (a host still on a v0.16 source resolves lts to v0.16.23 until it is moved up; see Upgrading). It closes two fleet-down split-brain classes seen on the v0.17.8 line. (1) Headless-cron token split-brain (#2469): a headless-cron reactive-quota 429 rotation routed through the shared lease-swap-or-defer verb but, on a defer_local / subprocess-error / malformed decision, fell through to a local claude-token rotate — overwriting the central token-updater's lease-selected token across the fleet in connected-authority mode (the 2026-07-12 token split-brain outage; the daemon already guarded this exact fall-through via #27006 Inc A, the cron-runner did not). The cron-runner now mirrors the daemon's connected-authority predicate (the same BRIDGE_UPDATER_AUTHORITY_ENABLED / break-glass gates + last-token-lease-ok marker hysteresis): in authority mode a defer / subprocess-error / bad-decision converts to a fail-closed authority_defer HOLD (audited, no local rotate), while disconnected / dark / break-glass keeps the classic local fallback byte-identical. (2) tmux socket split-brain containment (#2451, backport of #2450): closes the v0.17.8 rollout gap where a restart_agents=false layout-v2 reconcile could restore the named-socket daemon while legacy default-socket sessions were still alive — spawning a second fleet. Legacy exact-roster sessions now count as active, duplicate starts are suppressed, cross-socket split-brain is detected, and an explicit stop/restart drains only the exact managed sessions on both sockets before relaunching a single named-socket instance (no kill-server, no regex-wide selection, no shell command-template interpolation; the full automatic old-socket migration stays deferred to #2441). It carries the v0.17.8 tmux fixed-named-socket isolation (the session-wipe SPOF removal + usage-probe 403→401 fix + #2437 stabilization / #2438 picker backports), the v0.17.6 picker ghost-idle false-nudge fix, the v0.17.5 token-swap fleet-auth wedge + Teams silent reply-drop fixes, the v0.17.4 token-rotation durable-fix, and the v0.17.0 GA runtime forward unchanged. LTS_SERIES stays 0.17 (single converged v0.17.x line, no fork). See the v0.17.9 release notes for this patch, and the tagged v0.17.x CHANGELOG for the full v0.17.x detail (those v0.17.5–0.17.9 entries live on the release/0.17-lts line, not on main's CHANGELOG.md).

Upgrading: agent-bridge upgrade --apply (default --channel stable) resolves v0.17.9 as a single-atomic leap from any v0.7.x+ source — the v0.13.7–v0.13.10 cycle extracted the leap-path bodies to lib/upgrade-helpers/, closing the Bash 5.3.9 heredoc deadlock chain. A host already on a v0.17.x source advances with a plain agent-bridge upgrade --apply (a strict forward move; the downgrade guard is not engaged). Installs pinned to --channel lts land on v0.17.9 once they are on a v0.17.x source — the lts resolver reads the checked-out LTS_SERIES, which stays 0.17 (single converged v0.17.x line, no fork), so a host still on a v0.16.x source keeps resolving lts to v0.16.23 until it is moved up. That first move onto v0.17 is a one-time explicit agent-bridge upgrade --apply --ref v0.17.9 — bare --channel lts cannot self-advance a v0.16-tree host; after that hop, bare --channel lts stays on the v0.17 line. Migration is manual and per-host (there is no auto-upgrade — no host auto-moves), so stage a fleet canary → expand. See OPERATIONS.md §"Upgrade".