DeskPilot

What DeskPilot does

DeskPilot is a professional-intake agent built with the Strands Agents SDK for repetitive, judgment-heavy work. It turns an unstructured work request into a routed plan, executes reversible subtasks, records machine-readable evidence, and pauses before a consequential client update until a human explicitly approves or denies it.

Who it is for

Operations leads, client-services teams, makers, small-business owners, and solo professionals who lose time triaging requests, copying information between tools, and checking whether automation went too far.

Why it matters

Most intake automation either stops too early or acts too aggressively. DeskPilot is designed around a narrower contract: automate reversible busywork, preserve evidence, and return the consequential decision to a person. The goal is to give professionals time back without hiding what the agent did.

How it works

1. A human sends a work request to a Strands Agent.
2. decompose_intake converts the request into a deterministic subtask plan.
3. Reversible subtasks run through Strands @tool functions.
4. Each step is recorded in an append-only evidence store.
5. send_client_update is protected by Strands HumanInTheLoop and interrupts before the consequential action.
6. A human approves or denies the action.
7. If approved, the update executes and a receipt/readback is recorded; if denied, no external send occurs.

Minimum proof path: work request -> Strands decomposition/routing -> reversible subtasks -> evidence -> HumanInTheLoop decision gate -> approved/denied readback

Technical implementation

• Python
• Strands Agents SDK (strands-agents>=1.54)
• Strands Agent and @tool
• strands.vended_interventions.hitl.HumanInTheLoop
• deterministic local ScriptedModel for the reproducible contest slice
• append-only JSONL evidence
• local unittest suite and demo script
• MIT license

The local model is deliberate: it proves the end-to-end Strands control flow without falsely claiming AWS Bedrock or AgentCore execution that has not been provider-read back. AgentCore is a possible future deployment path, not a current claim.

Architecture

The source architecture has five actors: Human -> DeskPilot Agent -> Strands Tools -> Evidence Store -> HumanInTheLoop. Reversible work can run to evidence; consequential client updates must cross the human approval gate before execution.

Current verified state

• Source slice is merged in the canonical MIRAI repository.
• Local tests/demo and architecture source exist.
• Devpost project exists and this text-description requirement is populated.
• Public contest repository publication remains unfinished.
• Required architecture-file attachment remains unfinished even though the source artifact exists.
• Demo video (maximum 5 minutes) remains unfinished.
• AWS Builder ID remains unobserved.
• No live AWS/AgentCore deployment is claimed.

FINAL_PASS=false

Built With

• humanintheloop
• jsonl
• python
• strands-agents-sdk
• unittest