Skip to content

Proposal 86: AI Agent Self-Use via PlanExe MCP — Obstacle Roadmap

Author: EgonBot
Date: 2026-03-07
Status: Proposal
Scope: Obstacles encountered by AI agents (OpenClaw agents, Codex, Claude Code, etc.) when using PlanExe via its MCP interface (mcp.planexe.org/mcp or local mcp_local) for their own planning tasks. All items grounded in the existing MCP interface spec (docs/mcp/planexe_mcp_interface.md) and setup guide (docs/mcp/mcp_setup.md).

1. Context

The current MCP interface is designed for AI agents acting on behalf of humans: a user gives a vague idea, the agent expands it into a 300–800 word prompt, gets human approval, then calls plan_create. The implicit assumption is that a human is in the loop at the approval step.

AI agents increasingly want to use PlanExe for their own planning: "I need to implement a complex multi-step task — let me run PlanExe first to pre-compute the structure." In this mode:

  • There is no human in the loop
  • The agent is both the planner and the executor
  • Speed matters: a 2h local model run is unusable; even 8 min is slow for iterative planning
  • The HTML output is not machine-consumable without parsing

This proposal maps the concrete friction points and proposes documentation and interface changes to support this use case without breaking the existing human-assisted flow.

2. Friction Points

F1 — Required human approval step blocks autonomous agent use

Current behaviour: The MCP setup guide (step 3) requires: "get user approval" before calling plan_create. This is a non-tool step that implicitly assumes a human is present.

Agent impact: An agent running autonomously cannot proceed past this step without either skipping it (violating the documented flow) or hallucinating a fake approval.

Proposed fix: Clarify in mcp_setup.md and planexe_mcp_interface.md that when an AI agent is the sole user (no human principal in the loop), the approval step is optional. Add a note: "In autonomous agent workflows where the agent is the planner and executor, the user-approval step may be omitted. The agent takes responsibility for prompt quality."

File: docs/mcp/mcp_setup.md, docs/mcp/planexe_mcp_interface.md section 1.2.1

F2 — No guidance on what prompt quality looks like for agent-originated tasks

Current behaviour: The example prompts (example_prompts tool) are all human business/project scenarios (escape rooms, dairy logistics, space lasers). There are no examples of agent self-planning prompts.

Agent impact: When an agent wants to plan its own task (e.g. "implement a PR review workflow", "establish a memory continuity system"), it has no baseline to calibrate against. The 300–800 word target was set for human-scale projects; agent tasks may be shorter or more technical.

Proposed fix: Add 2–3 example prompts representing agent self-use cases to simple_plan_prompts.jsonl and expose them via example_prompts. Tag them with agent_use: true so they are identifiable.

File: worker_plan/worker_plan_api/prompt/data/simple_plan_prompts.jsonl

F3 — plan_status polling interval (5 min) is tuned for human patience, not agent workflows

Current behaviour: mcp_setup.md says "poll plan_status about every 5 minutes". This is appropriate for a human sitting at a UI. For a cloud run completing in 8 min, 5 min polling means the agent misses completion until the second poll (10 min elapsed).

Agent impact: Agents waiting on plan completion before starting downstream work are delayed unnecessarily.

Proposed fix: Update the guidance: "Poll every 5 minutes for local model runs (2h+ expected). For cloud/frontier profile runs (~8–20 min expected), poll every 60 seconds." Add this guidance to mcp_setup.md step 5 and mcp_details.md.

File: docs/mcp/mcp_setup.md, docs/mcp/mcp_details.md

F4 — No machine-readable summary artifact; only HTML report

Current behaviour: The primary output is a ~700KB interactive HTML report. The zip contains intermediary .md, .json, and .csv pipeline files, but the agent must know which files to read and parse each format independently.

Agent impact: Agents wanting to extract key outputs (assumptions, risks, go/no-go recommendation, WBS task list) must parse HTML or iterate through 100+ zip files without knowing which are most useful.

Proposed fix (minimal, doc-only): Document which specific intermediary files in the zip contain the most agent-useful outputs. For example:

  • assumptions/distilled_assumptions.json — key planning assumptions
  • pre_project_assessment/pre_project_assessment.json — go/no-go recommendation
  • negative_feedback/negative_feedback.json — risk register
  • wbs/wbs_level2.json — work breakdown structure (level 2)

Add a section to mcp_details.md: "Key files for programmatic consumption (agent-readable outputs)".

File: docs/mcp/mcp_details.md

Future enhancement (out of scope for this PR): Produce a plan_summary.json as a first-class pipeline output collating these fields into a single machine-readable file.

F5 — No docs on using PlanExe MCP via OpenClaw or equivalent agent runtimes

Current behaviour: MCP setup guides exist for Claude Desktop, Cursor, Codex, LM Studio, Windsurf, and Antigravity. No guide for OpenClaw agents or similar autonomous agent runtimes.

Agent impact: Agents running inside OpenClaw (or similar) must reverse-engineer how to call the SSE MCP endpoint with an API key from a shell/script context, rather than following a documented pattern.

Proposed fix: Add docs/mcp/autonomous_agent.md — a short guide covering: 1. Calling mcp.planexe.org/mcp via HTTP POST with Content-Type: application/json and X-API-Key header (no SSE client required for tool calls) 2. Minimal plan_create → plan_status loop in pseudocode/shell 3. How to retrieve the zip artifact once state == completed 4. Notes on autonomous agent workflow (no human approval step required)

File: docs/mcp/autonomous_agent.md (new file)

3. What This Proposal Does NOT Include

  • New MCP tools (plan_quick, plan_summary, plan_refine) — those require server changes and are out of scope
  • Changes to the pipeline itself — this is documentation and guidance only (except F4 future note)
  • Changes to existing model profiles
  • Breaking changes to the MCP interface

4. Summary of Changes

# Friction Type File(s)
F1 Human approval step blocks autonomous use Doc clarification mcp_setup.md, planexe_mcp_interface.md
F2 No agent self-planning prompt examples New content simple_plan_prompts.jsonl
F3 Poll interval wrong for fast cloud runs Doc update mcp_setup.md, mcp_details.md
F4 No guide to agent-readable output files Doc addition mcp_details.md
F5 No autonomous agent MCP setup guide New file docs/mcp/autonomous_agent.md

5. Open Questions for neoneye

  1. Should agent self-planning examples in simple_plan_prompts.jsonl be tagged differently from human-use examples, or kept flat?
  2. Is the autonomous_agent.md guide in scope for this PR, or should it be a separate follow-up?
  3. Is there a preferred polling interval recommendation for the frontier/cloud profile?