How mao-code/manta works
MANTA is an inference-time adaptive multi-agent system within an experiment and evaluation harness. The primary workflow configures a run, plans a task-specific topology, executes and audits collaboration turns, repairs the topology when needed, and records a final answer for benchmark evaluation. Across runs, process signals update a playbook; traces and evaluations feed descriptor analysis. The tree and excerpts do not expose benchmark-adapter source, so its internal wiring is left unspecified.
Main entry points: engine.py Topology engine · commands.py Run commands · runner.py MAS runner · main.py Experiment CLI
Components 19
Experiment entry
- Experiment CLI
main.py - Run commands
cli/commands.py - Experiment config
MAS/experiment.py
Adaptive MAS
- Topology engine
MAS/self_evolved/engine.py - Topology planner
MAS/self_evolved/planner.py - Trace auditor
MAS/self_evolved/auditor.py - Turn executor
MAS/self_evolved/executor.py - Topology spec
MAS/self_evolved/spec.py - Experience playbook
MAS/self_evolved/playbook.py
Collaboration runtime
- MAS runner
MAS/runner.py - Topology workflows
MAS/langgraph_engine.py - LLM client
MAS/llm.py
Evaluation and analysis
- Benchmark adapters
- Run artifacts
MAS/artifacts.py - Descriptor analysis
descriptor/experiment.py - Descriptor metrics
descriptor/metrics.py - Experiment outputs
cli/artifacts.py
Other components
- Experiment operator
- LLM provider
Connections 23
- Experiment operator → Experiment CLI: starts run
- Experiment CLI → Run commands: dispatches
- Run commands → Experiment config: loads settings
- Run commands → Benchmark adapters: selects tasks
- Run commands → MAS runner: runs tasks
- MAS runner → Topology engine: selects topology
- Topology engine → Experience playbook: loads experience
- Topology engine → Topology planner: requests plan
- Topology planner → Topology spec: proposes topology
- Topology engine → Turn executor: executes turn
- Turn executor → Topology workflows: runs collaboration
- Topology workflows → LLM client: requests completions
- LLM client → LLM provider: sends requests
- Topology engine → Trace auditor: audits trace
- Trace auditor → Topology engine: reports findings
- Topology engine → Topology spec: repairs topology
- Topology engine → Experience playbook: records experience
- MAS runner → Run artifacts: records trace
- MAS runner → Benchmark adapters: submits answer
- Benchmark adapters → MAS runner: returns evaluation
- Run commands → Descriptor analysis: analyzes runs
- Descriptor analysis → Descriptor metrics: computes metrics
- Run commands → Experiment outputs: writes results