Kuzushi Labs · Engineering notes

Engineering agent harnesses

An AI concierge and an operating system for hospitality. Eighteen articles on the custom runtime behind guest conversations and internal operations—from requirements and architecture to the tools, state, evaluations and controls that make the product operable.

Arunank Sharan · August 2026 · 18 articles · 22,131 words

assistant personas

2

technical chapters

18

architecture diagrams

18

The series follows a client platform built with Python, FastAPI, PostgreSQL and a Next.js operator console. It explains the patterns selected from OpenClaw and Hermes, the domain-specific booking and dispatch flows, and how capabilities are composed for each turn.

1. PART 01 · 8 min read

   From an AI concierge to an operations system

   Turning guest conversations and hospitality workflows into explicit requirements, authority boundaries and observable outcomes.

2. PART 02 · 7 min read

   Designing the harness: architecture, Hermes and OpenClaw

   The runtime boundaries, the selected open-source patterns and the custom control plane built around hospitality operations.

3. PART 03 · 7 min read

   The run contract: orchestration, lifecycle and idempotency

   Following one message from authenticated ingress to a durable result, with duplicate delivery, cancellation and retries handled explicitly.

4. PART 04 · 7 min read

   Skills and on-the-fly composition

   How two assistants assemble the right instructions, model and tool surface for each turn without becoming a swarm of agents.

5. PART 05 · 7 min read

   Prompt assembly is an engineering pipeline

   Stable personas, fresh domain evidence, parallel hydration, prompt ownership and cache-aware composition.

6. PART 06 · 7 min read

   Memory: what was said, what was shown and what is true

   Separating conversation history, property knowledge and durable booking or issue state, with scope and freshness preserved.

7. PART 07 · 6 min read

   The model transport and the bounded agent loop

   Provider adapters, streamed tool arguments, cancellation, repair passes and the distinction between generated and deliverable text.

8. PART 08 · 6 min read

   Tools as data, execution as a controlled boundary

   Compiling tool definitions into handlers, validating backend contracts and keeping model arguments inside authenticated scope.

9. PART 09 · 7 min read

   Authority, approvals and the trust boundary

   Separating conversational confirmation from authorization, scoping actions and designing durable human approval.

10. PART 10 · 6 min read

    Booking: let the model converse and the system decide

    Search, calendar verification, stable options, fresh checkout links and evidence-backed claims across multiple turns.

11. PART 11 · 6 min read

    From a complaint to work that someone owns

    Issue intake, deduplication, staff availability, dispatch ladders, escalation and honest status updates.

12. PART 12 · 6 min read

    One event spine, several ways to deliver the result

    Typed runtime events, SSE and synchronous projections, UI components, audit delivery and the boundary between generation and notification.

13. PART 13 · 6 min read

    The control plane: changing behaviour without losing control

    Versioned prompts, tools, skills, agents and models; ownership, hot reload, deployability and reproducible releases.

14. PART 14 · 6 min read

    Evaluation starts with the operational scenario

    A practical evaluation system for multi-turn booking, issue handling, tool correctness and release decisions.

15. PART 15 · 6 min read

    Judges, human review and regression evidence

    Groundedness versus helpfulness, evidence packets, paired comparisons, position bias and human calibration.

16. PART 16 · 7 min read

    The operator console is part of the agent product

    An inspectable path from a guest complaint to routing, prompt assembly, tool evidence, cost and a new regression case.

17. PART 17 · 7 min read

    Latency, concurrency and the cost of a completed task

    Critical-path analysis, nested limits, cache-aware accounting and load tests that reflect real agent behaviour.

18. PART 18 · 7 min read

    Reliability, release discipline and operational value

    Crash boundaries, outbox delivery, configuration consistency, deployment checks and the next steps toward a reusable harness.

Read the client platform case study, explore the diagram gallery, or continue with financial RAG and conversational AI.