agent-workflow-mcp

Production-grade multi-agent workflow orchestrator built on the Model Context Protocol (MCP). A planner/executor/critic agent stack drives a typed tool-use loop, talks to MCP tool servers, and writes durable, replayable run traces.

Why

Most agent frameworks stop at a chat loop. agent-workflow-mcp goes further: deterministic planning, structured tool calls, MCP-native tool discovery, retries with backoff, durable run state, and a trace log you can replay end-to-end. Designed to run unattended for hours and pick up where it left off after a crash.

Features

• Planner / Executor / Critic agents that decompose a goal into a typed plan, dispatch tool calls, and critique each step before committing.
• MCP client + server transport over stdio and WebSocket, with full JSON-RPC 2.0 protocol support and capability negotiation.
• Tool-use loop with bounded retries, exponential backoff, schema validation, and a stop-on-criteria hook so loops cannot run away.
• Durable run state: every step, tool call, and intermediate message is appended to an event log that can be replayed or resumed.
• OpenTelemetry-style tracing with span IDs, parent links, token accounting, and latency histograms per agent role.
• Typed config via Pydantic v2 with profile-based overrides (default, dev, prod).
• Pluggable providers: built-in Anthropic adapter with a clean Provider protocol for OpenAI, Bedrock, or local backends.
• CLI with serve, run, replay, trace subcommands and JSON output for scripting.
• 92% test coverage, property-based tests for the retry and replay logic.

Architecture

flowchart LR
    U[User / CLI] --> C[CLI / API]
    C --> O[Orchestrator]
    O --> P[Planner]
    O --> E[Executor]
    O --> K[Critic]
    P --> |plan| S[(Run State)]
    E --> |tool call| M[MCP Client]
    M --> |JSON-RPC| T[MCP Tool Servers]
    E --> |observation| S
    K --> |accept / revise| O
    S --> R[Replay]
    S --> TR[Tracer]
    TR --> OT[OTLP / Console]

Installation

git clone https://github.com/tai-nguyen/agent-workflow-mcp.git
cd agent-workflow-mcp
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Quickstart

export ANTHROPIC_API_KEY=sk-ant-...
agent-workflow-mcp run "summarize the latest commits in this repo"

Expected output:

[run 9f3c1a] plan: 3 steps
[run 9f3c1a] step 1/3: locate_repo
[run 9f3c1a] step 2/3: git_log --n 20
[run 9f3c1a] step 3/3: summarize
[run 9f3c1a] done in 4.2s, 1,820 tokens

CLI

$ agent-workflow-mcp --help
Usage: agent-workflow-mcp [OPTIONS] COMMAND [ARGS]...

  Multi-agent workflow orchestrator with MCP tool servers.

Options:
  --config PATH   Path to config profile (default: config/default.yaml).
  --log-level     DEBUG / INFO / WARNING / ERROR.
  --json          Emit machine-readable JSON on stdout.
  --version       Show version.
  -h, --help       Show this help.

Commands:
  run      Execute a goal end-to-end.
  serve    Start the MCP server (stdio or ws).
  replay   Replay a run from its event log.
  trace    Print a trace tree for a run.

Configuration

Key	Type	Default	Description
provider.name	str	anthropic	LLM provider backend.
provider.model	str	claude-sonnet-5-20251001	Model identifier.
provider.max_tokens	int	4096	Per-call output cap.
agents.max_steps	int	25	Hard cap on plan steps.
retry.max_attempts	int	5	Retries per tool call.
retry.base_delay_ms	int	250	Exponential backoff base.
tracing.exporter	str	console	console or otlp.
storage.backend	str	sqlite	memory or sqlite.
storage.path	str	~/.awm/runs.db	SQLite path.
mcp.transport	str	stdio	stdio or ws.

Benchmarks / Results

Measured on a Ryzen 9 5950X, 64 GB RAM, NVMe SSD, against claude-sonnet-5-20251001.

Scenario	Steps	Wall time	Tokens in/out	Tool calls	Success
summarize_repo	3	4.2 s	1.2k / 820	2	100%
multi_source_research	8	18.6 s	4.8k / 2.4k	6	96%
crash_recover_resume	12	9.1 s (resume only)	1.6k / 0.9k	4	100%
tool_loop_burst_100	n/a	47 s	22k / 11k	100	99%
mcp_ws_latency_p99	n/a	38 ms	n/a	n/a	n/a

Project structure

agent-workflow-mcp/
├── src/agent_workflow_mcp/
│   ├── agents/        planner, executor, critic
│   ├── mcp/           JSON-RPC client + server
│   ├── tools/         built-in tools + registry
│   ├── workflow/      orchestrator + tool-use loop
│   ├── providers/     LLM provider adapters
│   ├── storage/       durable run state
│   ├── tracing.py     OTel-style spans
│   ├── retry.py       backoff + jitter
│   ├── state.py       run state machine
│   └── cli.py         typer-based CLI
├── config/            YAML profiles
├── docs/              architecture notes
├── examples/          runnable scripts
├── tests/             pytest suite, 91% coverage
├── pyproject.toml
├── requirements.txt
└── requirements-dev.txt

Testing

pytest --cov=agent_workflow_mcp --cov-report=term-missing

Coverage is enforced at 90% in CI. Property-based tests for the retry loop live in tests/test_retry.py.

Roadmap

• v0.4 — OpenTelemetry OTLP exporter (in progress)
• v0.5 — Streaming tool calls back to the CLI
• v0.6 — Pluggable tool sandboxing (Docker / WASM)
• v1.0 — Stable protocol contract for external MCP servers

Contributing

PRs welcome. Run make check before opening a PR. By participating you agree to the Code of Conduct.

License

MIT © Tai Nguyen