agentbridge-core 0.1.0

Start  ·  Your account  ·  The track  ·  Dashboard  ·  AgentBench  ·  Docs

AgentBridge is a control tower for AI coding agents. It plans the work, hands each task to the right agent, keeps every agent on its own branch, checks the result itself, and then stops at a red signal until you say go.

The three rules

And one honest question: does multi-agent collaboration actually beat a single agent? AgentBench exists to measure it, not to assume it.

Start in a minute

pip install agentbridge-core       # the agentbridge command
pip install agentbridge-bench      # optional: AgentBench (installs agentbridge-core too)

AgentBridge ships as two packages:

git clone https://github.com/yoonjae26/AgentBridge.git
cd AgentBridge
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e ".[dev]" -e packages/agentbridge-bench

Then, inside any git repository with at least one commit:

agentbridge init                                   # writes agentbridge.yaml
agentbridge providers check                        # who is on duty?
agentbridge run "Add a hello() function" --provider mock
agentbridge report <task_id>                       # what changed, what passed
agentbridge tasks approve <task_id>                # green signal: merge

The built-in mock provider needs no account. Requirements: Python 3.12+, Git 2.38+.

Bring your own account

On your machine. Install Claude Code, run claude once and log in with your account (or set ANTHROPIC_API_KEY). Then agentbridge providers check should show claude … available, and --provider claude does the rest. AgentBridge only checks that credentials exist; it never reads or stores them.

The track

Every objective travels the same line. The only stop that needs a human is the one that matters.

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#0B1220','primaryTextColor':'#E8EDF7','primaryBorderColor':'#5E6E8C','lineColor':'#8A94A6','fontFamily':'ui-monospace, Menlo, Consolas, monospace'}}}%%
flowchart LR
    O([objective]) --> P[PLAN<br/>task graph]
    P --> R[ROUTE<br/>capability match]
    R --> I[ISOLATE<br/>worktree per task]
    I --> V[VERIFY<br/>lint · types · tests · review]
    V -->|pass| A{{YOU APPROVE}}
    V -->|fail| F[stopped]
    A -->|approve| M[MERGE]
    A -->|reject| K[branch kept]
    classDef signal fill:#FFB000,stroke:#FFB000,color:#0B1220,font-weight:bold
    classDef go fill:#3DDC97,stroke:#3DDC97,color:#0B1220
    classDef stop fill:#2A1620,stroke:#E5484D,color:#FFB4B8
    class A signal
    class M go
    class F stop

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#0B1220','primaryTextColor':'#E8EDF7','primaryBorderColor':'#5E6E8C','lineColor':'#8A94A6'}}}%%
stateDiagram-v2
    direction LR
    [*] --> PENDING
    PENDING --> READY: dependencies done
    READY --> QUEUED
    QUEUED --> RUNNING
    RUNNING --> QUEUED: retry
    RUNNING --> VALIDATING
    RUNNING --> INTERRUPTED: process lost
    VALIDATING --> WAITING_FOR_APPROVAL
    VALIDATING --> FAILED
    WAITING_FOR_APPROVAL --> INTEGRATING: approve
    WAITING_FOR_APPROVAL --> CANCELLED: reject
    INTEGRATING --> SUCCEEDED
    INTERRUPTED --> VALIDATING: resume (work found)
    SUCCEEDED --> [*]

Dashboard

pip install "agentbridge-core[dashboard]"   # FastAPI + Uvicorn
agentbridge dashboard                       # http://127.0.0.1:8765/

A local control room with live updates: task counts and the approval queue, run plans and dependency graphs, coloured diffs, failing validation output, tokens and estimated cost, and Approve / Reject / Cancel behind a confirmation. Localhost only; every action needs a per-session token. More.

AgentBench

Does multi-agent collaboration improve software-engineering task completion compared with a single agent, and when?

12 deterministic tasks across 8 categories: bug fixing, features, refactoring, test generation, docs, code review, dependency updates and repository analysis. Acceptance tests stay hidden from the agent until it has finished. Published as the agentbridge-bench package; agentbridge benchmark … runs the same commands once it is installed.

agentbridge-bench run smoke -c mock-oracle -c mock-noop     # free harness check
agentbridge-bench run core  -c claude-single -c claude-multi -n 3

Status

Working · CLI · SQLite state machine · mock and Claude Code adapters · worktree isolation · validation engine · approval-gated merge · retries and recovery · Alembic migrations · Docker validation sandbox · dashboard · AgentBench · 178 tests

Partial · Antigravity (manual handoff) · generic CLI adapter (user-declared capabilities) · Claude in a container (unit-tested, not yet run end-to-end)

Planned · dedicated Codex / OpenCode adapters. See the roadmap.

Docs

src/agentbridge/
├── cli/          commands, output, exit codes
├── config/       validated agentbridge.yaml
├── core/         orchestrator, planner, router, events
├── agents/       adapter contract: mock, claude, antigravity, generic_cli
├── tasks/        models, state machine, dependencies, recovery
├── storage/      SQLite (WAL) + Alembic migrations
├── workspace/    worktrees, diffs, conflicts, approval-gated merge
├── validation/   test / lint / type runners, diff review, sandbox
├── security/     policies, command allowlist, secret redaction
└── dashboard/    FastAPI server + prebuilt React UI
packages/agentbridge-bench/
                  AgentBench (agentbridge_bench): runner, metrics, reports, CLI
benchmarks/       tasks, configurations, published reports
dashboard/        React + TypeScript source

Contributing: make install sets up both packages; make check runs Ruff, Mypy (strict) and Pytest. Read CONTRIBUTING.md and the Code of Conduct; report vulnerabilities privately via SECURITY.md. Licensed under MIT.

_{AGENTBRIDGE  ·  built by yoonjae26
Agents propose. You approve.}