Selected work

Shipped systems, shown by the hard part.

Autonomous departments, agent products, on-device systems — live and in real use.

See shipped systemsStart a conversation

What we mean by harness

The tool belt around the model.

Founders do not need a smarter demo. They need an AI system that can touch real workflows without corrupting state, skipping approvals, leaking data, or running up cost in silence.

Agent operating system

The model is the engine. The harness is everything that lets it work.

Agent

Tool belt

Control

State

Judgment

Visibility

Recovery

A prompt can answer. A harness can safely read, write, remember, recover, and explain its work.

Tool belt

Tools agents can actually use

MCP servers, plugins, APIs, native app control, and internal workflow integrations.

Control

Permissioned action

Read-versus-write boundaries, scoped credentials, approval gates, and audit trails.

State

Memory and state

Checkpoints, shared files, durable context, and exact procedural memory where embeddings are wrong.

Judgment

Evals and model routing

The right model for the task, with validators, routing rules, and fine-tuning only where it earns its keep.

Visibility

Observability and cost control

Traces, per-output cost, retry budgets, SLIs, alerts, and the ability to explain what happened.

Recovery

Recovery paths

Idempotency, circuit breakers, resumable workflows, and failure modes designed before launch.

Featured shipped system

A whole agent department, not a chatbot.

The point is not that there are agents. The point is that they have tools, state, gates, traces, and recovery paths.

AI department

An autonomous AI content department

One brief becomes finished, scheduled, multi-platform content through a governed agent tree.

Fourteen named specialists are only the first layer. The workflow fans into depth-4 leaf agents for research, drafting, QA, packaging, routing, and publishing. The durable piece is the harness: shared file bus, checkpointed state, idempotent dispatch, eval-gated routing, cost tracing, and human approval gates.

What it proves

  • One orchestrator governing a depth-4 agent tree
  • Fourteen core specialist lanes fan out into leaf workers
  • Shared file bus with checkpointed state and idempotent re-runs
  • Eval-gated model routing, cost tracing, and on-prem observability
  • Human approval gates before anything publishes
14
core specialist lanes before the tree fans out
Depth 4
leaf agents below the top-level specialists
Orchestrator dispatches over a Shared file bus to depth-4 agent tree (extract, route, draft, review, publish, and more), whose output passes through a human approval gate to Approved output; gated by checkpointed dispatch and eval-gated routing.
Orchestratorcoordinator
Dispatch / routing
  • checkpointed dispatch
  • eval-gated routing
Shared file bus
  • extract
  • route
  • draft
  • review
  • publish
  • …

depth-4 agent tree

Human approvalgate
Approved outputscheduled + published

What this prevents

The expensive failures are quiet.

CEOs do not buy checkpoints. They buy confidence that an AI workflow will not rerun a payment, skip approval, lose state, or become impossible to debug when volume rises.

A comparison of two agent designs. Left, the prompt-first chain: Tool runs, then Worker dies, then Workflow retries, then Business state corrupts — This is how a working demo becomes an operations problem.. Right, the harness chain: Checkpoint written, then Bounded tool runs, then Failure is traced, then System resumes, which loops back to its checkpoint — The harness makes failure recoverable by construction..

Prompt + tool call

Tool runs

email, payment, row, ticket

Worker dies

after the side effect lands

Workflow retries

without knowing what happened

Business state corrupts

quietly, at production speed

This is how a working demo becomes an operations problem.

Harnessed agent

Checkpoint written

state before side effects

Bounded tool runs

permissioned + idempotent

skip duplicate

Failure is traced

retry budget and cost guard

trace

System resumes

no duplicate action

resume

The harness makes failure recoverable by construction.

  • MCP servers
  • Plugins
  • Permissions
  • Evals
  • Checkpoints
  • Observability
  • cost ok

tool_call = agent_id + step + payload_hash

Case studies

More systems in production.

AI product · web + desktop

A conversational AI agent product

A natural-language agent that drives a native real-time engine, in paying users' hands.

A full-stack AI product built on a single-coordinator, multi-tool agent: a conversational interface wired to a native real-time engine, with subscriptions, cloud infrastructure, and end-to-end observability. Shipped across web and desktop, in production.

What it proves

  • Single-coordinator, multi-tool conversational agent
  • Native real-time engine (C++ DSP) under the chat layer
  • 12+ external API integrations
  • Subscriptions and a credit system on Stripe
  • Web + desktop across five platforms, fully observable
On-device · compliance-first

An on-device capture system for regulated data

Clinical-grade transcription where the audio never leaves the machine.

A HIPAA-adjacent capture system, compliance from day one. Transcription runs on-device with local ML, so sensitive audio never leaves the machine — a custom low-latency audio driver, macOS System Extensions and XPC, a sandboxed, crash-isolated build. Shipped.

What it proves

  • On-device ML transcription — no audio leaves the machine
  • Custom audio driver, sub-10ms latency
  • macOS System Extensions + XPC, sandboxed and crash-isolated
  • Audit logging from day one
AI pipeline

A document → structured-data pipeline

Messy specs in, schema-validated data out — every field citing its source page.

Spec, quote, and proposal PDFs and DOCX in; schema-validated fields out, each citing the source page. Layout-aware parsing preserves tables, and output is enforced through the tool-use API — no fragile JSON parsing — then handed to an editable review UI with Excel and PDF export.

What it proves

  • Schema-validated output enforced via the tool-use API
  • Per-field source-page citations
  • Layout-aware parsing that preserves tables
  • Editable review UI with Excel / PDF export

Bring us in when

Where founders use us.

These are buying moments, not service categories. They usually happen after the demo works and before the system is safe enough to become part of the business.

Buying moment

You want an AI department

A whole function should run from one business input, with agents decomposing the work and humans approving the output.

  • Agent orchestration
  • Shared state bus
  • Approval gates

Buying moment

Your agent is brittle

It works until an API times out, a tool runs twice, costs spike, or nobody can explain why the workflow made a decision.

  • Checkpoints
  • Idempotency
  • Traceability

Buying moment

AI needs to be the product

The model needs a real product around it: auth, billing, data, APIs, native integrations, background jobs, admin surfaces, and deployment.

  • Full-stack SaaS
  • Native systems
  • Payments

Buying moment

Documents need to become decisions

PDFs, audio, specs, quotes, and operational documents need to become schema-valid outputs your team can inspect, cite, approve, and export.

  • Extraction
  • Citations
  • Review workflows

Open-source infrastructure

Tools other engineers run.

MCP servers and agent tooling we build in the open. Public, installable, used outside our work.

reaper-mcp

An MCP server giving AI agents live control of a target application. FastMCP, on PyPI.

  • 58tools across 9 categories
  • Installable from PyPI

goalkeeper

An open-source Claude Code plugin for contract-driven, multi-agent goal execution. A subagent judge gates completion against an explicit Definition of Done, so a passing validator can never ship a placeholder.

  • v0.3.0
  • 8skills
  • 80lifecycle test assertions

clawd-cursor

A 40-tool desktop-automation MCP server (TypeScript, macOS). We contribute rather than author it — landing macOS stability, GPU detection, and accessibility fixes upstream.

  • 40-tool MCP server
  • Upstream contributor: macOS stability, a11y

OpenClaw Live Events plugin

An OpenClaw plugin, published on ClawHub, that lets agents search live events through the Ticketmaster Discovery API.

  • Published on ClawHub
  • Ticketmaster Discovery API
  • OpenClaw community plugin

OpenClaw OpenMontage plugin

An OpenClaw plugin, published on ClawHub, that teaches agents to drive AI video production — explainers, trailers, motion graphics — through a local OpenMontage pipeline.

  • Published on ClawHub
  • AI video production
  • OpenClaw community plugin

Where this work has gone

Industries served.

Real proposals and shipped engagements.

  • Real estate & proptech
  • Construction & AEC
  • Healthcare (HIPAA-adjacent)
  • Creator economy & marketing
  • E-commerce & retail
  • Fintech
  • Logistics & trucking
  • Hospitality
  • Podcasting
  • Robotics
  • ML research

Want the same rigor on your system?

We will tell you honestly whether we are the right team.

Start a conversation