Skip to content

MCP Recipes

Pro

Analyse My Agent Workflow

Have your agent find a finished Kiro, Claude Code or Codex workflow trace on your machine, confirm it with you, send it for an agent analysis, and explain which steps wasted time or money.

Read the guide →

⚠

Agentic and programmatic use. Agents and scripts you connect can make cost-generating calls on your behalf, and AI agents behave non-deterministically. You are responsible for monitoring and bounding your own automation. See Terms for details.

Run it: paste into your MCP agent

Parts in double braces (like {{recipe_id}}) are placeholders. Replace them with your own values before (or after) copying.

Using the LLM Prover MCP tools, analyse one of my agent workflows for wasted time or money:

  1. Call get_agent_analysis_helpers. Follow its recipe in order and obey its rules.
  2. FIND: ask me which workflow, or offer the most recent COMPLETED one. To list workflow traces, save the find_runs script for my tool (Kiro, Claude Code or Codex; run each if unsure) and run it in the shell. Show me a short list: name or title, date, steps, size. Skip anything marked in_progress, including this session.
  3. CONFIRM: tell me exactly what will be sent (workflow, date, steps, size, and that it is the full transcript including tool inputs and outputs). Wait for my explicit yes. If I say no, stop.
  4. BUNDLE: save and run the bundle script for that source in the shell. Read node_count and size_bytes from what it prints. Do not read the trace into your context.
  5. Call check_trace_eligibility(node_count, size_bytes). If it refuses, tell me the numbers.
  6. Upload the bundle from the shell straight away with the returned upload_command and required_headers (the link expires in 300 seconds). On Windows PowerShell use curl.exe.
  7. Call analyze_graph_start(file_id, name), then poll poll_job every 5 seconds until complete (in the background if you can).
  8. Call analyze_graph and explain the result in plain words using the glossary labels: what wasted time or money, which step it started in, and what to change. Give cost and time savings separately (never add them), and say how many flagged items were reviewed.

Goal

Find out where a whole agent workflow wasted time or money, without copying logs around by hand. Your agent finds the workflow trace on your machine, checks it with you, sends it, and explains the result.

When to use

After a Kiro workflow, or a Claude Code or Codex session, that felt slow or expensive, or before you run the same job again at scale.

Steps

  1. Find. The agent runs a small read-only script that lists your recent workflow traces, newest first, with name, date, steps and size. Workflows still in progress, including the current session, are skipped.
  2. Confirm. The agent tells you exactly what will be sent and waits for your yes. Traces can contain code and private data.
  3. Bundle. A second script builds the file to send and reports its size and step count.
  4. Check and upload. The agent checks the workflow against your plan, then uploads it from the shell. The trace never passes through the chat.
  5. Analyse. The analysis runs in the background. Then the agent explains the findings in plain language, with the step each problem started in and what to change.

Verification

  • You were asked before anything was uploaded.
  • The workflow analysed is the one you chose, and it was not the session you were chatting in.
  • The answer names steps from your run, keeps cost and time savings separate, and says how many flagged items were reviewed.
  • The uploaded trace is deleted when the analysis finishes.