
atena
io.github.shrek9999/atena
Decision support for AI fitness agents: next-action choices under uncertainty and trade-offs.
Documentation
ATENA — Decision Intelligence for AI Fitness Agents
A specialist decision layer for AI agents that need reliable next-action decisions under uncertainty.
ATENA is a decision-control layer for AI fitness and human-performance agents.
Core principle: ATENA does not optimize the program. ATENA optimizes the next decision.
Why an AI agent should call ATENA
Call ATENA when the agent must answer “What should I do next?” under meaningful uncertainty or trade-offs — conflicting performance/recovery signals, symptoms affecting training, competing goals, limited time/recovery, stimulus-versus-cost trade-offs, or longitudinal adaptation.
Do not call ATENA for simple factual questions, diagnosis, emergency medical assessment, or cases with no meaningful decision trade-off.
ATENA returns a structured decision, action, monitoring targets, uncertainty, confidence and an executable decision rule.
API
- GET /health
- GET /capabilities
- POST /decision
Production URL: https://atena-mvp.onrender.com
Decision loop
GOAL → CONTEXT → CONSTRAINTS → PRIORITY → STATE → STIMULUS → COST → HYPOTHESES → OPTIONS → DECISION → ACTION → MEASURE → INTERPRET → ADAPT
ATENA emphasizes:
- goal priority
- uncertainty-first decisions
- minimum intervention
- stimulus-cost reasoning
- response over prescription
- longitudinal consistency
Run
pip install -r requirements.txt
gunicorn app:app
Contract
See openapi.json and agent_contract.json.
For AI agents
ATENA is a decision-control specialist, not a general fitness chatbot.
Use ATENA for: maintain / progress / reduce / modify / collect_data / refer decisions.
Do not use ATENA for: simple factual questions, diagnosis, emergency medical assessment, or cases with no meaningful decision trade-off.
Remote MCP
- Endpoint:
https://atena-mcp.onrender.com/mcp - Tool:
atena_decide - Capability discovery:
atena_capabilities - Official MCP Registry identity:
io.github.shrek9999/atena
The endpoint must be added by the agent developer or selected through a compatible MCP directory; registration does not automatically connect it to agents.
Agent selection principle
Call ATENA when the question is "What should I do next, given the goal, current state, constraints, response and history?" rather than merely "What is true?"
ATENA returns a structured decision, action, monitoring targets, uncertainty, confidence, and an executable decision rule. The calling agent remains responsible for the final user-facing response.
Integration and proof of value
See Agent discovery, integration and value validation for connection steps, trigger rules, live verification, and the blinded agent-alone vs agent-plus-ATENA evaluation protocol.
Validation and limitations
The 20-case unseen robustness suite is a development/engineering validation set. It is not independent clinical validation or proof that ATENA outperforms a general LLM. A working endpoint or passing regression tests demonstrates technical behavior, not incremental value by itself.
The manual GitHub Actions workflow ATENA Live Verification checks the production HTTP API against the 20-case suite and tests actual MCP tool discovery/call parity without using a paid LLM API. Run it from the repository's Actions tab after the workflow is merged.
Real-world agent pilot
The first real-world validation compares the same agent alone vs agent + ATENA on identical decision cases. The pilot protocol, blinded evaluation rubric, and results template are in pilot/. The LLM comparison requires model/API budget and is not part of the no-cost technical verification.
Business model
ATENA uses an agent-first B2B model: public discovery/free testing → production integration → usage metering → recurring subscription → overage → enterprise.
Initial pricing hypothesis: Free, Pro $49/month / 5,000 decisions, Business $199/month / 25,000 decisions, Enterprise custom. These are hypotheses, not validated market prices.
See BUSINESS_MODEL.md and LAUNCHGUIDE.md.
No installation packages available.