christiangrey922/multi-agent-workflow-lab

cd multi-agent-workflow-lab && pnpm install --frozen-lockfile
pnpm demo

Usage

• •Run pnpm demo to see a deterministic review chain (orchestrator → researcher → analyst → reviewer → evaluator).
• •Run example workflows with pnpm mawl run workflows/02-parallel-research.yaml (negative examples exit non-zero when blocked).
• •Inspect a run: pnpm mawl trace show <run-id>, pnpm mawl evaluate <run-id>, pnpm mawl graph <run-id>.
• •Replay with pnpm mawl replay <run-id> dry-run (or exact / model-rerun / tool-rerun, which refuses external side effects without explicit permission).

Key features

• ✓Delegation engine with parent/child task graphs, target/capability checks, depth and fan-out limits, and loop detection.
• ✓MCP client adapters (official stdio and Streamable HTTP) plus an in-memory mock connector; MCP data is treated as untrusted.
• ✓Nine deterministic evaluation rules, eight diagnostic dimensions, run comparison, and an optional model judge.
• ✓Versioned prompts (semver + SHA-256, provenance and trust classification), deny-by-default permissions, redaction, and budget controls.
• ✓Security-testing scenarios: prompt injection, escalation, malicious MCP/tool output, secret leakage, traversal, recursion, and resource exhaustion.

Best for

Teams evaluating, comparing, and policy-testing multi-agent / MCP workflows in a deterministic lab before production.

Caveats

• ⚠Experimental v0.1.0 release candidate — suitable for local development, evaluation, and policy testing; not production-hardened infrastructure.
• ⚠The restricted local sandbox is not equivalent to hardened container/VM/microVM isolation, and network denial is best-effort.
• ⚠No claims of being prompt-injection-proof, sandbox-escape-proof, or fully zero-trust.
• ⚠The default path uses MockModelProvider; real models require an OpenAI-compatible provider with credentials injected explicitly (the CLI intentionally won't auto-load them).
• ⚠Requires Node.js 22+ and pnpm 11.