Task Harness: Structured Task Management for Long-Running Agents

Where this fits

You're three sessions deep into a feature request. The agent rebuilt the auth flow, refactored the database layer, and somewhere along the way the original scope got fuzzy. You need a single source of truth that survives context resets and session handoffs. Task Harness solves this by converting a high-level requirement into a structured feature_list.json plus four harness files, giving every future agent session a clear map of what's done, what's pending, and how to verify each piece.

Why agents benefit

• Persistent state beyond context windows – The JSON task list is written to disk, so a new agent session can pick up exactly where the last one stopped without re-reading the entire conversation history.
• Explicit verification criteria – Each feature includes a verification field, so the agent can self-check its work against concrete conditions instead of guessing at "done."
• Priority-aware execution – With priority and category fields, the agent can sequence work logically (foundation before components) and skip low-priority items when time is short.
• Cross-session dependency tracking – Harness files encode the relationships between features, letting the agent detect blocking issues (e.g., a missing foundation feature) before starting dependent work.
• Human-readable progress reporting – Because the task list is plain JSON, you can diff it across sessions to see exactly what changed, without parsing log noise.

Practical scenarios

Multi-session feature build

You ask for a "user dashboard with charts and export." The agent analyzes your repo, breaks it into 12 features across categories (foundation, layout, components, api), and writes the harness files. Session one builds the foundation; session two (hours later) reads the JSON, sees what's incomplete, and continues without re-asking.

Regression tracking across sprints

Your team adds new functionality while refactoring old code. Task Harness keeps a passes flag per feature. When a later session modifies a file, it can run the stored verification steps to confirm nothing broke, and mark features as failed if needed.

Onboarding a new agent to an existing project

A fresh agent instance needs to contribute to a codebase it has never seen. Instead of a vague prompt, point it at the harness files. It reads the task list, understands the architecture from the file paths, and starts executing the highest-priority unverified feature immediately.

Add it to your agent workflow

Install the skill directly from the command line:

npx skills add kangarooking/kangarooking-skills --skill task-harness

Then invoke it with a project name and requirement:

[task-harness] "ecommerce-store" "Add a product search with filters and a cart page"

The agent will analyze your repository, generate feature_list.json and the associated harness files, and start executing the first feature. For more context on how this fits into the broader agent-skill ecosystem, see the OpenAgentSkill listing.

Compare before adopting

Before committing, evaluate task-harness against alternatives like simple TODO files or manual checklists. Compare on: quality signals – does the generated structure actually reduce agent confusion? Maintenance freshness – is the repo active and does it track Anthropic's latest harness recommendations? Alternatives – some projects may prefer a lighter-weight approach like a single tasks.md; others need the full JSON schema. Workflow fit – does your team already use structured project management? If so, the JSON format might integrate better than a Markdown list.

Why it is worth tracking

Task Harness has 568 stars and is actively maintained by kangarooking, with a clear methodology rooted in Anthropic's effective harnesses research. The skill is designed for teams that run long-lived agent processes and need deterministic state. It's worth evaluating when your agent projects start spanning multiple sessions or when you notice context drift causing rework. Because it's installable via npx skills, you can test it on a small feature branch without affecting your main workflow. Track it if you're building serious multi-session agent systems, not just single-shot prompts.