Enterprise AI agent platform
Say it once.
Watch it become a working system.
Millwright is an AI agent that turns your words into a living workflow — building any missing tool along the way. What it hands back keeps working for you: same quality, every time, as many times as you need.
This is the product's real flow. The agent designs and verifies freely — and the door to the real world opens only when you say so. Creativity up front, certainty where it counts.
From a conversation to a working system
Missing a tool? The agent builds it for you, on the spot.
Three screens tell the whole story — each one mirrors the real Studio handling a real request.
read my new email, summarize it in chinese, and send the summary to my slack
slack.messenger (tool) — your workspace doesn't have this yet. The agent can build it for you, and you'll see the code before anything is installed.
- package.manifest-and-install
- lifecycle.start-health
- lifecycle.shutdown
Built-in peace of mind
There's a reason you can hand it the keys.
| candidate | revision_6f9fac7923f9d041026d49f0 |
| source | sha256:3aa47ac13731…3294bc |
| authority | 3 operations · exact resources, exact digests |
| budget | 10 requests · 64,000 in / 16,384 out tokens · 300 s |
| grant | single-use · held by the platform, never the browser |
| receipt | approval-receipt_77683e9edef11cb459aa71cf |
- Four eyes, by design.Requesting and approving are two different seats — a healthy double-check is built into the flow, even for a team of one.
- Exactly what you approved runs.Approval is tied to the precise version you reviewed. If anything changes, the platform simply asks you again.
- Budgets you can relax about.Spending limits are enforced by the platform itself — no run can quietly exceed what you agreed to.
The foundation of trust
A living map of everything your agent can do.
Every capability — built-in, official, or agent-built — is a verified node on one capability graph. The agent works only from this map, so you always know exactly what it can touch.
- Real capabilities only.Workflows can reference only what actually exists on the map — an imagined tool is caught long before it could ever reach production.
- The map grows with you.Every new capability earns its place: generated, tested in a sandbox, and installed with your confirmation.
- Approvals remember the map.Each approval keeps a snapshot of the map it was judged on. If the world changes underneath, the platform asks you again — automatically.
Every run tells its story
Look back at any run, any time.
Not a chat log — a complete, replayable record. Even when something goes wrong, you get evidence you can act on, not a mystery.
A real record from a real run — deep research on a user's topic, completed within budget, result verified.
A third way
Freedom where it creates. Certainty where it counts.
Static workflow builders
Everything by hand
You assemble every node yourself, and a missing connector means waiting for someone else to ship it.
Free-running agents
Brilliant, but fleeting
Impressive once, different every time — nothing your team can rely on, reuse, or audit tomorrow.
Millwright
Create freely, run reliably
The agent's creativity goes into building — designing workflows, writing new tools. What you run is a system you trust: reviewed, approved, repeatable.
Imagination, set free. Execution, made certain.
Inventory → Market
What the agent builds today becomes your team's assets tomorrow.
Every add-on, workflow, and result is born versioned, verified, and carrying its own history — exactly what it takes to be shared with confidence. Your local inventory exists today; a place to publish and exchange is where we're headed.
Add-on
generated/slack-messenger · v1.0.0
sha256:384a7303…
Built by the agent, proven in a sandbox, installed with a verified fingerprint.
Workflow
email_summary_to_slack · r4
revision_6f9fac79… + approval receipt
A working system with its full story attached — and it travels: export it as a portable, verified bundle.
Run result
run output · sha256:a260f1e4…
Every result stays linked to the exact run, version, and approval that produced it.
A marketplace for working systems
Publish your systems and tools — history and verification included — to your whole organization, or to a wider market. What one team builds, every team can trust, because its story travels with it.
An honest roadmap
What works today, and what we're building next.
We'd rather show you real machinery than beautiful promises.
Working today
- Agent-drafted workflows — typed, compiler-verified, semantically tested
- Self-built tools — generated, sandbox-proven, installed on your confirmation
- Two-seat approval — exact versions, real budget limits, durable receipts
- Guarded runs — single-use grants, a complete event record, verified results
- Secrets stay server-side — credentials and inputs never reach the browser
- Local-first Studio — your machine, your data, private by default
- Export machine bundle — your system, its tools, and its history in one verified, portable file
- Results in the browser — successful runs show their verified output right in the Studio
Coming next
- Hosted publish & deploy — run your systems as managed endpoints; the seam is already in the Studio
- Artifact preview & download — file artifacts are metadata-only in the browser for now, by design
- Deeper tool hardening — even finer permissions for agent-written add-ons
- Terminal UI — the same experience, from your shell
- Hosted inventory & market — share verified assets across your org, and beyond
Start with one real piece of work.
Bring one task you actually do. Watch the agent build the system, watch it pass your approval, and watch it run — live, in front of you.
Book a demo cognitech.dev