TaskPilot

TaskPilot

TaskPilot is a Professional Agents project for the Agents for Humans Hackathon. It handles repetitive professional preparation in the background, completes safe reversible work autonomously, and surfaces only decisions that could create a meaningful external consequence.

Problem

Knowledge workers lose time to repetitive coordination: turning notes into next steps, preparing follow-ups, organizing tasks, and deciding whether an action is safe to execute automatically. Most assistants either stop at chat or act too aggressively.

TaskPilot uses a simple contract: prepare automatically, pause before consequences.

How it works

1. A user provides a work task and optional context.
2. A Strands Agent plans the work.
3. Safe preparatory work is drafted automatically.
4. A deterministic policy tool classifies any proposed external or state-changing action.
5. Low-risk reversible work can continue automatically.
6. External communication, payments, purchases, signatures, deletions, legal/privacy-sensitive actions, and other consequential changes are surfaced for human approval.
7. The agent returns an audit-style summary and never claims an action occurred unless a tool actually performed it.

Technical approach

The project uses the Strands Agents SDK for agent orchestration and tool use. The reference model path uses a local Ollama model, allowing the project to run without paid AWS runtime usage. The repository includes deterministic approval-gate tests and an architecture diagram.

Why it matters

The strongest background agents should remove repetitive work without forcing people to supervise every harmless step. At the same time, they should not silently cross financial, legal, privacy, communication, or irreversible boundaries. TaskPilot makes that boundary explicit and inspectable.

Track

Professional Agents.

Built With

• ai-agents
• ollama
• open-source
• python
• strands-agents-sdk