Handoff

An autonomous maintenance-coordination agent for property managers — built with the Strands Agents SDK on Amazon Bedrock.

🔴 Live demo: https://0fmmk8vbt0.execute-api.us-east-2.amazonaws.com/ — submit a maintenance scenario (try midnight_flood with after hours checked), watch the agent triage → quote vendors → pause for your approval on big spend, then act as vendor and tenant to complete the cycle. State is durable; come back later and the board remembers.

The problem. Property managers spend 3+ hours every day on vendor coordination — the #1 time drain in the profession (NAA 2025). A single work order takes 8–15 manual touches: triage, vendor calls, quote chasing, scheduling, status checks, invoice matching. 14% of vendor dispatches no-show; 64% of tenants get zero proactive updates while they wait. Maintenance — not rent — is the top driver of negative reviews and non-renewals.

The agent. Handoff owns every handoff in that chain. Tenants report an issue; Handoff classifies severity, discovers prices from the vendor bench, dispatches the best-fit vendor with a complete job card, keeps the tenant informed at every step, chases stalled jobs on a schedule, and matches invoices against authorized scope. It runs in the background and only surfaces for real decisions — spend above policy threshold, after-hours emergencies, low-confidence triage, invoice discrepancies.

Architecture

flowchart TB
    subgraph People["People"]
        J[Judge / Visitor]
        P[Property Manager]
    end

    subgraph Edge["Public Surface"]
        AGW[API Gateway HTTP API]
        FN[Lambda · handoff-dashboard<br/>FastAPI via Lambda Web Adapter]
    end

    subgraph Brain["Agent Layer — Strands Agents SDK"]
        CA[Coordinator Agent<br/>Strands agent loop + tools]
        SE[SafetyEnsembleProvider<br/>hazard-keyword escalation over LLM]
    end

    subgraph AWS["AWS us-east-2"]
        AC[Bedrock AgentCore Runtime<br/>same brain, invoke API]
        BB[Amazon Bedrock]
        DB[(DynamoDB · shared durable state)]
        EB[EventBridge Scheduler<br/>nightly sweep]
    end

    J -->|browser| AGW --> FN
    P -->|browser| AGW
    FN <-->|read/write state| DB
    FN -->|invoke| AC
    AC -.->|hosts| CA
    EB -->|sweep| AC
    CA <-->|inference| BB
    CA --> SE

Loading

Two deployment surfaces share one durable store: the Lambda-backed dashboard you can open in a browser, and a Bedrock AgentCore Runtime exposing the same agent as an invocation API (new_request / decide / sweep / status) for EventBridge-driven background sweeps. Approve a gate from one session; the ticket's dispatched state shows up everywhere else.

Design principles

1. Probabilistic reasoning, deterministic mechanics. LLM decisions never touch ticket state directly. Every side effect goes through idempotency-keyed tools backed by atomic (update_ticket) storage — retries after crashes or approval waits can never double-dispatch or double-text.
2. Durable human gates. Above-threshold and after-hours dispatches persist the exact intended vendor+price before pausing. The PM's approval — minutes or days later — resumes precisely that action.
3. Escalation as capability. Low-confidence triage, universal vendor declines, and invoice discrepancies route to a visible human queue instead of failing silently.
4. Eval-gated judgment. Triage accuracy is measured against a 22-case library before anything ships (see docs/research/aug21-summary.md: live-model prompt optimization 64% → 100%, stable, via an automated propose→eval→keep/discard loop).

Quick start

uv venv --python 3.13 .venv
uv pip install -e ".[dev]"
.venv/bin/python -m pytest tests/          # 53 tests: reliability core + redteam pins
.venv/bin/python -m handoff.demo           # headless demo run
.venv/bin/python -m uvicorn handoff.web.app:app --port 8731   # dashboard

Without AWS credentials the app runs on deterministic providers (same interfaces, same tests). Set HANDOFF_MODEL_PROVIDER=bedrock (+ AWS_PROFILE) to run the Strands coordinator on Amazon Bedrock — see .env.example.

License

MIT — see LICENSE.