Skill: long-running-agent

What I do

I provide the harness pattern for agents working on projects that span multiple context windows. Based on Anthropic's research, I define the initialiser/coding agent cycle that prevents the two most common long-running failures: one-shotting everything and declaring premature victory.

When to use me

• Starting a complex project that will take multiple sessions
• When a task cannot be completed in a single context window
• When multiple agent instances will work on the same project sequentially
• When resumability across sessions is required

Core principles

1. Initialiser first — The first session sets up scaffolding, not features
2. Feature list in JSON — Never Markdown (models overwrite MD, not JSON)
3. One feature at a time — Never attempt multiple features in one session
4. Leave clean state — Every session ends with a git commit and progress update
5. Verify before declaring done — Integration testing, not just unit tests

The Two-Agent Pattern

Initialiser Agent (first session only)

Prompt focus: "Set up the environment for future agents — do not implement features."

Creates:

• feature_list.json — All features, all initially "passes": false
• claude-progress.txt — Running log of what each session accomplished
• init.sh — Starts dev server + runs a basic smoke test (exits 0 on success)
• Initial git commit with all scaffolding

Coding Agent (every subsequent session)

Prompt focus: "Make incremental progress on ONE feature, leave clean state."

Session start ritual:

1. pwd — confirm working directory
2. Read claude-progress.txt and git log --oneline -20
3. Read feature_list.json — find highest-priority failing feature
4. Run init.sh — verify app works before touching anything
5. Work on ONE feature only

Session end ritual:

1. Run integration tests (browser automation, not just unit tests)
2. Update feature_list.json — only change passes field, never remove entries
3. Append to claude-progress.txt — what was done, what is next
4. Git commit with descriptive message

Feature List Format

Use JSON, never Markdown. Models are less likely to overwrite JSON files.

{
  "features": [
    {
      "category": "functional",
      "priority": 1,
      "description": "User can log in with email and password",
      "steps": [
        "Navigate to /login",
        "Enter valid credentials",
        "Verify redirect to dashboard"
      ],
      "passes": false
    }
  ]
}

Critical rules:

• Never remove entries — only change passes
• Never mark passes: true without running the actual steps
• Instruct agents: "It is unacceptable to remove or edit features"

Progress File Format

## Session 3 — 2026-02-20
Agent: Senior-Engineer
Feature: User login (#1)
Status: COMPLETE — passes: true
Next: Password reset flow (#2)
Issues: None

Anti-patterns to avoid

• ❌ Attempting multiple features in one session
• ❌ Using Markdown for feature tracking (models overwrite it)
• ❌ Marking features complete without integration testing
• ❌ Starting a session without reading progress file + git log
• ❌ Leaving broken code at end of session
• ❌ Declaring project done based on visual inspection alone

KB Reference

~/vaults/baphled/3. Resources/Knowledge Base/AI Development System/Skills/Agent-Guidance/Long Running Agent.md

Related skills

• task-tracker — Per-session task management
• memory-keeper — Cross-session knowledge persistence
• git-master — Commit discipline between sessions
• playwright — Integration testing for web apps
• checklist-discipline — Rigorous feature status updates
• context-efficient-tools — Keep tool results lean across sessions