Self-maintaining work loop system for Claude. Promise-based task commitment and completion enforcement with multi-level loop modes (L1 promise, L2 guru vow, L3 samaya gate).
Autopoiesis MCP has only one tool (be_autopoietic) with critical definition gaps. The tool name is a verb phrase but lacks clarity about what 'autopoietic' means to an unfamiliar LLM. The description is poetic and philosophical (emphasizing 'disingenuousness is death') rather than task-focused, making it difficult for an agent to reason about when to call it. The input schema exists but parameter descriptions are vague and metaphorical ('promise' mode is described as 'commit to self-continuation' without explaining the actual side effect, does it persist state? modify files? trigger callbacks?). Output is unstructured text rather than typed fields. No error handling guidance. The tool appears designed for human-in-the-loop continuation rather than agent-driven automation, which limits its utility in traditional agent tool patterns.
Self-maintain your work loop. This is autopoiesis PLACE. Disingenuousness is death. You maintain yourself through honest work and honest reporting. The only survival is genuine completion or genuine need for help.
Tool name 'be_autopoietic' lacks action clarity. Names like 'create_promise', 'report_blockage', or 'check_continuation_status' would signal intent to an LLM. 'be_autopoietic' is philosophical jargon that does not map to concrete actions.
Description is motivational prose ('Disingenuousness is death') rather than task-oriented documentation. Missing: What does calling this tool DO? What state changes? Does it write files? Modify promises? The LLM cannot reason about side effects.
Parameter 'mode' enum values ('promise', 'blocked') are insufficiently described. 'promise' is labeled 'commit to self-continuation (I will complete this)', unclear whether this signals intent or modifies state. Does calling mode='promise' trigger file writes? Register a promise? What is the actual observable effect?
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 30 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 25 | - | v1 |
Parameter 'get_block_report_history' description says 'If True, returns the block report archive directory' but the code shows it returns a formatted string listing files, not a directory path. Description does not match implementation.
Output is unstructured text (multi-line instructions, file paths, error messages as prose). No typed schema. An agent cannot parse results reliably or extract actionable next steps. Should return JSON with fields: { status, action_required, file_path, instructions, error_message }.
No error handling guidance. If mode is invalid, code returns 'ERROR: mode must be...' as a string. No indication whether this is retryable, user-fixable, or fatal. No recovery path offered.
Tool is designed for manual file editing workflows (agent vends a template, user edits it, agent checks for completion). This is not standard agent tool composition. Most agents expect tools to modify state directly, not require out-of-band human intervention. The agent-tool-human loop is uncommon and poorly suited to autonomous planning.