Put a workforce of AI agents on your real production work.
It compounds. Every run makes the next one faster, better, and cheaper.
Legate runs a workforce of governed AI agents on your real work. Capability compounds and carries across teams, so the work gets better and cheaper as you scale. And because every agent runs inside a bounded mandate with a tamper-evident record, the proof an auditor wants comes for free.
Two ways to put AI on real work. Both break. There is a third.
Companies are moving real production work onto AI agents now. The two common approaches do not hold up. One ships work nobody checked. The other never compounds.
Vibe-coding ships work nobody checked.
Prompt, accept, ship. Fast for a demo. It is unbounded and unauditable. It does not survive production, and it does not survive an audit.
A coding harness re-pays its tax every time.
A harness gives one team reusable execution. Useful. But it does not compound. Every new team, project, and model upgrade re-pays the setup. Your standards and know-how do not carry.
Legate compounds.
Every validated run is codified back into shared, versioned know-how. The next run starts higher. Capability carries across teams and forks. You own the machine, not just the outputs.
Why buy this year: your organization stays accountable for what its agents do. "The AI did it" is not a defense. The EU AI Act high-risk duties land Q4 2026 to Q2 2027. Legate gives you the proof, on your schedule. Production is why you win. The window is why you sign this year.
Three ways in. Adopt one. Get full value. Stay there.
What is the human for?
As the agents take on the production volume, your people move up. They do not disappear. They move to higher-leverage work: building the capability, administering the platform, operating the business, and designing the self-learning organization itself. The open standard underneath Legate, CARE, asks this exact question by name. The three lanes below are the answer.
Each lane is a distinct person with a distinct job and a standalone payoff. Nothing forces you up the stack.
Builder: author a capability once, and every project can use it.
- Who: an engineer who builds reusable capability. Engine code with its agent, skills, and rules, as one deliverable.
- Standalone value: install the open SDK and build on a shipped engine. No platform and no consultant required.
- Payoff: what you build once becomes leverage every downstream product can draw on. A workaround built once becomes ecosystem-wide.
- Maturity: engine and open SDK shipped. Apache-2.0, on PyPI today.
How we build: the shipped engine you build on is kailash. You build on it. You do not rebuild it.
Administrator: make improvements spread without drift.
- Who: the platform engineer who runs the distribution machine and the shared method and standards.
- Standalone value: run the distributor as pure knowledge management. Classify what is shared vs local, scrub disclosures, and distribute. No product code originates here.
- Payoff: every project inherits improvements on its next pull. No per-project re-decision. The reason the loop compounds instead of fragmenting.
- Maturity: distribution and classification shipped. A read-only control center is alpha. The write and actuation path is roadmap.
How we run: csq is our open Foundation session substrate that runs the agent fleets around the clock. Apache-2.0. It is how we dogfood, not a product you buy. loom-command is our internal control room over the foundry. Internal tooling, alpha, not a product.
Operator: put already-onboarded agents on real business work.
- Who: a business user who directs governed agents in plain language and runs the six-phase workflow.
- Standalone value: pull the capability you need on your own cadence. Build products. No obligation to contribute anything back.
- Payoff: stand up a governed agent workforce for real outcomes. Your validated work can feed back to builders and lift the whole base.
- Maturity: shipped and in use, including one live client fork in production.
Two applications you can run on their own today.
Legate is not one monolithic product. These run standalone. They also compose (see the loop).
Aegis: run a governed agent workforce at organization scale.
- What: a typed org tree of agents, each inside a bounded mandate, with a tamper-evident audit trail.
- Standalone: runs in your own environment. On-prem, air-gapped, or sovereign. No other part of the stack required.
- Why it is a wedge: deploy-anywhere is something a free single-database tool structurally cannot serve.
- Maturity: shipped and in production, including one live client fork. Audit signing ships but stays off until you provision keys. The hash-chain is always on.
Envoy: give one person a single delegated agent at the edge.
- What: a personal edge agent carrying signed authority rooted in one human. Bounded grants, hash-chained ledger, key recovery, skill ingest.
- Standalone: local-first, self-hosted, free forever under Apache-2.0. Never gated by the commercial tier.
- Maturity: code complete and tested. Zero pilots. We say working alpha, not production. The Rust speed path is not shipped, so we do not claim a speed number.
Build capability. Put it to work. Feed what it learns back. Repeat.
This is the mechanism behind "it compounds." It runs in both directions.
How we build and run at this velocity: csq (our open session substrate, Apache-2.0) runs the agent fleets around the clock. kailash is the shipped engine underneath. loom-command is our internal control room over the foundry, alpha. These are how we produce. They are not products you buy.
Compose the lanes into a self-sufficient cognitive enterprise. Roadmap / direction of travel
Run all three lanes together and each function becomes a workspace whose know-how lives in versioned files and compounds across the company. Compose enough functions and the competence lives outside any individual's head. You can fork the whole ecosystem and run independently.
The proof an auditor wants falls out for free.
You do not configure governance as a feature. The work already runs inside bounded mandates and a tamper-evident record, so the proof is a by-product.
Every agent runs inside a bounded mandate.
Five dimensions: money, actions, time, data, and communication. A delegated mandate can only ever get narrower. Access fails closed. The default answer is deny.
The trust lineage is signed.
Delegations are cryptographically signed. Decisions run on a four-level gradient. Anything unrecognized is held for a human. It fails toward a person, not toward action.
The audit trail is tamper-evident.
Every governed action is written to an append-only, hash-linked chain that verifies itself. Alter a record after the fact and verification breaks. This is the record an auditor reads.
One engine, two SDKs that agree.
A proprietary high-performance build and the open SDK produce the same governance result via a shared canonical format. A regulated buyer is never locked to one binary to trust the result.
Governed against open standards a neutral foundation owns. Not our private rulebook.
Legate conforms to open standards for agent trust, governance, and accountability. Integrum does not own them. The Terrene Foundation publishes them under CC BY 4.0, structurally entrenched against capture, including against Integrum. The reference Python SDK is open source and Foundation-owned. Anyone can build on the standards, including a funded competitor. That is by design. A neutral standard cannot be quietly changed to suit us.
Terrene Foundation owns
the open standards (CC BY 4.0) and the open SDK (Apache 2.0). Anyone may build on these.
Integrum conforms
builds the commercial engine and products. Proprietary. Conforms to the open standards.
A published standard is not a shipped enforcement engine. Do not read spec maturity as product maturity.
The engine runs today. The platform and the appliance are the roadmap.
Running today
- The compiled engine and its governance layers.
- The open SDK on PyPI. Install it now.
- Aegis in production, including one live client fork in the client's own secured environment.
- The within-ecosystem compounding cascade.
On the roadmap
- The integrated platform that wires the separate governed surfaces into one product.
- Usage-based metering.
- The sovereign hardware appliance. Air-gapped, pre-loaded, kept current by a one-way pull. The edge primitive runs today. The box is a forward build.
- Envoy Fleet, the org-scale tier.
- Aegis go-to-market packaging and SOC 2 Type II.
- Cross-ecosystem fork sync.
We mark our own roadmap. That is how you know the "today" column is real.
Check us directly. No logo wall.
The engine is inspectable.
Install the open SDK from PyPI. The governance behavior is in the code.
The standards are open.
Read them. They are published by a neutral foundation under a Creative Commons license. Conform to them yourself if you like.
The depth is production-proven.
One live client fork runs today in the client's own secured environment. We describe deployments by category, not by customer. Confidentiality is theirs to give, not ours to spend.
A team that marks its own roadmap is a team you can trust on what it marks done.
Legate is built by Integrum.
Legate and the engine beneath it are built by Integrum. Integrum builds commercial products in production, then pledges the open-substrate work to the neutral Terrene Foundation under a public commitment. The open standards and the open SDK belong to the Foundation, not to Integrum.
Integrum is led by its founder, Dr. Jack Hong, as CEO and Chief Architect. He is the face and the technical mind behind the platform. This is a permanent role by design.
Three ways to engage. Pick your lane.
For the builder
The open SDK is on PyPI and the standards are open. Build, conform, probe.
For the operator and the enterprise
Talk to us about a governed agent workforce in your own environment.