Autonomous Agentic Operating System exposing agent process management, cost control, governance, and multi-agent workflow orchestration as MCP tools
Server OS exposes 9 tools with complete input schemas and descriptions. Naming follows verb_noun convention (list_, create_, get_, run_). Descriptions are detailed (100-250 chars) and include prerequisites and side effects. However, output schemas are not formally documented, responses are inferred from code (model_dump calls). Error handling is minimal: run_task returns error objects but lacks recovery guidance. Parameters lack enums and format constraints (e.g., budget_usd unbounded, capabilities free-form string array). No pagination support despite tools returning lists. Security: no secrets exposed as params, but no permission gates or scope declarations visible. Composition is sound, tools chain well (create_agent → run_task → get_agent).
Create an agent process with intent, USD budget, and capabilities. Returns agent_id for later run_task / get_agent calls. Does not execute a goal. Side effect: writes the process table. Default model is mock-gpt.
Create specialist agents and run a sequential multi-agent workflow. Default roles are planner then worker, sharing budget_usd. Use when a goal needs more than one agent. Do not use to run a single existing agent (run_task). Side effects: creates agents + tasks and marks the workflow completed.
Get full state of one agent process by id. Use after create_agent or list_agents. Read-only. Missing ids return an error object. Does not run work (run_task).
Return governance allow/deny records from the policy engine. Use after a blocked run_task. Read-only. Not the dollar ledger (get_cost_ledger).
Return the token/$ cost ledger for this process. Use to audit spend after run_task. Read-only. Not get_audit_log or get_metrics.
Output schemas not formally documented. Tools return model_dump(mode='json') but no schema definitions visible in code. LLMs cannot plan downstream calls without knowing response structure.
Parameters lack enums and format constraints. 'capabilities' accepts free-form string array with no validation. 'budget_usd' unbounded (could be negative or 1e9). 'model' free-form string. LLMs will hallucinate invalid values.
Error handling lacks recovery guidance. run_task returns {'error': '...'} but does not tell LLM what to do next (retry? lookup? ask user?). No error categorization (retryable vs fatal).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Return runtime counters (agents created, tasks). Use for health checks. Read-only. Does not include cost rows or policy events.
List in-process agent records (id, status, spend, budget, capabilities). Use before run_task to discover agent_id values. Read-only snapshot. Does not create agents (create_agent) or start work (run_task).
List capability names that can be granted to agents. Use before create_agent to choose the capabilities argument. Read-only.
Run one goal on an existing agent until completion, policy deny, or budget stop. Blocks up to 120s. Missing agent_id returns an error object. Use create_agent first. Not for multi-agent orchestration (create_workflow). Side effects: task row + spend + save().
No pagination support. list_agents and list_available_tools return unbounded lists. Large result sets will blow context window. No limit, offset, or cursor parameters.
No permission gates or scope declarations. Tools like create_agent and run_task modify state but lack visibility into who can call them or what permissions are required. No audit trail visible in tool definitions.