From expert knowledge to a verified AI agent. Days, not months.
You write what the agent must do and must never do. Assize compiles that spec into a working, reproducible agent and proves it against a rubric before it ships.
The studio works today and is free while in beta. The open format and reference CLI are Apache-2.0, public at launch.
- Rule of Two≤ 2 of 3 powers
- Must-nots3 boundaries set
- Mandatory budgetscost + action caps
- Groundingcite or refuse
- Evidence rulesaudit trail on
Software has compilers. Agents don't.
Software has compilers. Code is written, compiled, and tested, and its behavior is known before it ships. Building an AI agent has no such workflow.
Every company building agents today starts from scratch. Domain experts describe what the agent should do, engineers translate that into prompts and code by hand, then spend weeks or months testing edge cases until the agent is trusted enough to deploy. There is no compiler from a specification to a reliable agent, and no systematic way to prove an agent behaves as intended.
The market already prices the gap. Teams hire RAG and agent engineers, scarce and expensive, to hand-build what is almost always the same first agent: one that answers with a company's knowledge and acts under rules. The demo is easy. Proving it behaves, and never misbehaves, is the work that takes months.
- 6+ months
- to put one agent in production
- 4 to 12 months
- per custom agent
- US$150k to 300k
- to hand-build one, estimated
The goal is simple: building a reliable agent should feel like compiling software. Deterministic, repeatable, and backed by tests instead of trial and error.
Expert, compile, verify.
Three steps turn a specification into a working agent with proof behind it, the same way source turns into a signed build.
- 01
Expert
Describe what the agent must and must never do: goals, tools, permissions, spend limits, behavior rules. In a visual editor or from the CLI. Your experts' documents go in, and the agent answers with retrieval and cited sources.
- 02
Compile
The platform validates that spec against a written standard, so mistakes surface while you design instead of in production, and turns it into a working, reproducible agent.
- 03
Verify
Automatic evals prove the agent meets its rubric before it ships, and every release leaves signed, reproducible evidence. The verdict is plain: SHIP or BLOCKED.
Use it from Claude over MCP, from any other LLM over the CLI, and as a check in GitHub Actions with no API keys in CI.
Built for the people who ship agents.
- AI agencies and consultancies shipping agents for clients
- In-house engineering teams building internal or customer-facing agents
- Engineers who want spec as code, a CLI and CI/CD, not another chat box
- Not a no-code builder yet, that layer comes later on the same engine
- Not a host or a runtime that operates your agent
- Not an eval dashboard bolted on at the end
- Not a framework rival, LangChain, LangGraph and CrewAI are integrations
Everyone else builds or measures. Assize does both, in one pipeline.
Assize compiles a complex agent from a declarative spec and verifies it in the same pipeline. The agent is data you version, review and reproduce, not code you hand-craft. No agent ships without passing its rubric.
Open core.
Open
Free · Apache-2.0
The spec format and the reference CLI. Works with any LLM. Public at launch.
- Spec format
- Reference CLI
- Any LLM
Studio
Per seat, at launch
The product: visual builder, compiler, verification and eval history. Bring your own API key. Free while in beta.
- Visual builder + compiler
- Verification + eval history
- Bring your own API key
Enterprise
Talk to us
For larger teams and regulated work: the versioned compliance packs and close onboarding, priced per engagement.
- Versioned compliance packs
- Per vertical
Pricing is per seat, not per agent and not per token. No number here yet, still being validated with the first teams.
The agent release checklist.
Fifteen questions a governed agent should answer before it ships, plus a scorecard to grade one you already have. The same standard the gate runs, on paper. No email required to read it.
An assize was the old sitting where a standard was set and a thing was measured against it: the assize of bread, the assize of cloth. That is the job. An agent faces a written standard and ships only with evidence it was met.
assize, noun · a session of measure and judgment
Tell me what you are building.
The studio is open, you can start now. If you would rather talk before you dive in, leave your details and you will hear back.
Got it.
Thank you. You will hear back soon.