Hallucination
The agent asserts what it was never told and could not know.
Intent Studio · The work layer
Your people and their agents, working from the same intent.
An intent is one piece of work — a line changeover, a migration, an audit response — written down as something a team and its agents can both work against. Intent Studio is where that gets built: questions, context and trade-offs, until it is sharp enough to act on.
Why agents fail
Agents don't fail because the model is weak. They fail because the intent was thin.
Given a vague brief, an agent does the worst thing possible: it carries on. It fills every gap with a guess, quietly and confidently. Each guess becomes the basis for the next step. By the time you see any output, it is built on guesses nobody agreed to.
Iteration
Round after round of "not quite, try again" — each one a full paid run.
Rework
The most expensive people in the building fixing an agent's guesses.
Corrupted plans
Work that was wrong at step two and only visibly wrong at step twenty.
You cannot see the gap while it is still cheap to fix. You can see it clearly once it is not.
What it does
An intent is a record, not a thread.
One written intent
Sections that follow the work, so nothing important gets assumed. The agent adds the ones this piece needs. Most intents start here:
At a glance · Goal · In scope · Out of scope · Definition of done · If it breaks · Systems affected · Why now
Claim states
Every line is marked Proposed, Agreed, Contested or Escalated, with a name against it. Disagreement stays visible instead of being smoothed over, and an agent cannot approve its own work.
Proposed
Agreed
Contested
Escalated
The autonomy dial
You decide how much the agents settle alone.
Five levels, set per workspace, per type of work, per risk level and per agent. You can lower it at any time; raising it needs someone to approve. So you can let agents decide more, and still defend having let them.
| Level | What it may do alone | What that requires |
|---|---|---|
| Propose | Suggest claims, questions and evidence. A person agrees to everything. | Nothing. This is the safe floor. |
| Settle the immaterial | Close low-consequence items alone; escalate anything material. | That materiality is defined by policy, not by the model's judgement. |
| Settle within a domain | Act unsupervised inside an intent type where its measured record supports it. | A track record that is measured, and specific to that domain. |
| Build and prepare | Produce artifacts and ready them for release without a human in each step. | That the gates still hold — attestation and approval unchanged. |
| Act, with gates | Carry the work through, stopping only where policy demands a named human. | Halt and revoke are instant, and every step is reconstructable. |
Holds at every setting: attributable · grounded · bounded · revocable
How it runs
On our servers, or on yours.
Use it hosted by us, or run it on your own infrastructure. Either way the workspace, the intents and the record of decisions stay together, over the models you choose — with Precepta underneath or without it. It does not need anything else to be useful.
Who it's for
Any team coordinating people and agents.
Built for people and agents working together on the same job. It helps most where teams who do not share the same language have to agree — a plant, a supply chain, controls and quality, all on one change.
How it is licensed
Contributors, not headcount.
The people who review, approve and check are the reason the record is worth having. They should never be the reason it costs more.
What you get
The workspace, today.
One written intent whose sections follow the work, claim states with attribution, the five-level autonomy dial, and role-based collaboration — over the models you choose, wherever you run it.
See it on one of your own intents.
The quickest way to understand it is to run something you are actually working on through it. Half an hour, your work, no slides.
Your team's decisions are now on the record. Precepta records what actually happened when the agents ran.