The sovereign orchestrator — connects all MCP AI suite libraries into an autonomous agent
26 tools with mixed quality. Naming is generally verb-first and clear (run_task, get_task_status, list_tasks). Descriptions present but often terse (10-50 chars), below the 50-200 char LLM-optimized baseline. Input schemas visible and properly typed for all tools, but many lack parameter-level descriptions. Output schemas not documented. Error handling and recovery guidance absent. No tool annotations (readOnlyHint, destructiveHint). Composition is reasonable, tools chain via task_id and namespace, but descriptions don't explain when to use each tool vs. similar ones (e.g., run_task vs. run_background vs. spawn_agent).
A/B test two constitutions on a goal: runs each, reports success/cost/token deltas + a winner.
Cancel a scheduled job.
Abort a running task
Side-by-side comparison of two task runs: goals, turns, tokens, cost, tools used, success.
Create and run a multi-agent TaskForce. Agents collaborate on a goal using a pattern (sequential, parallel, supervisor, debate). Use templates for quick setup.
Aggregated metrics: tasks completed/failed, total tokens, avg tokens per task, avg duration, top tools and agents by usage.
Descriptions too terse (avg 35 chars, target 50-200). Most tools lack context for LLM selection. E.g., 'Check task progress' (20 chars) doesn't explain when to call get_task_status vs. get_turns vs. get_trace.
No parameter-level descriptions. 'namespace' appears in 8+ tools but is never explained. 'mode' in run_task/run_background lacks guidance on when to pick 'react' vs 'ltp' vs 'hybrid'.
No output schemas documented. LLMs cannot plan downstream calls or extract fields. E.g., run_task returns what? A task_id? Full task object? Status? Latency?
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
View current constitution rules (3 layers: baseline, meta-improvement, user)
Check the status of a background operation.
Check task progress
Get timeline of turns for a task with timing info (timestamps, durations, tools used).
Get ReAct turn history
Kernel health + server connections
View kernel task audit log (task completions, failures, costs)
View kernel configuration
Token costs, latency, model usage
List available sub-agent types
List all background operations.
List all scheduled jobs.
List available TaskForce templates with their agent configurations.
List all tasks
Run a task in the background. Returns immediately with an operation ID. The task executes asynchronously — poll with get_operation or wait for a notification.
Submit a task for autonomous execution. Uses LTP (compile-once, deterministic) by default in hybrid mode, falls back to ReAct if needed.
Schedule a task for later or recurring execution. Supports: once (delay/at), cron, interval, watch.
Update user rules (layer 3). Does not affect meta-improvement rules (layer 2).
Spawn a specialized sub-agent for a task.
Fire a webhook event
No error handling or recovery guidance. Tools like cancel_task, set_constitution, trigger_webhook are destructive but lack confirmation patterns or error messages to guide LLM recovery.
No tool annotations. 26 tools lack readOnlyHint, destructiveHint, or idempotentHint. LLMs cannot infer which tools are safe to retry or which modify state.