Kohentra

Agents

Business agents that have to earn your trust.

Kohentra calls in specialised agents when their expertise counts. Each one is designed, versioned, qualified on your own judgements, and gets autonomy only if you grant it.

Illustration
Agent · Ticket classification
VersionStatusIn your workspace
1.0.0 In service Qualified on your judgements
1.1.0 Under evaluation Same tickets, no effect · disagreements to judge
1.2.0 Prepared Not published: runs nowhere

A published version never changes. None inherits the trust earned by the previous one.

The right expertise, at the right time

A lone agent is not Kohentra. Kohentra decides which agent looks at what, with what context, and when a piece of work has become pointless.

  1. Detecta situation in your activity.
  2. Gather the contexthistory, linked objects, conclusions already known.
  3. Decide whether expertise helpsand launch nothing if it doesn't.
  4. Pick the capabilityand open up only its scope.
  5. Collect the conclusionwith its evidence, kept exactly as produced.
  6. Never redowhat is already solved or already in progress.

Designed, not hard-coded

Describe the need the way you would explain it to a colleague: the agent designer helps you turn it into a complete definition that you review before anything is created. You can also start from an existing agent.

Kohentra also provides its own agents, and you choose which of their versions you run. No new version replaces yours without your decision.

  • Its domain

    and the resource it works on, chosen among those Kohentra actually knows in your workspace.

  • What it examines

    the situations in scope, and what it is trying to determine.

  • Its conclusions

    and what each one means: this is what you will judge it against.

  • Its reasoning

    rules first; a language model only when needed, and bounded.

  • Its context

    the related objects it may read, if Kohentra can establish the relation on your data.

  • Its actions

    the ones it may propose. Declaring them does not authorise them.

An agent starts with no authority at all

Every new agent, and every new version, starts in observation: it analyses your real data and produces real conclusions, but nothing changes in Dolibarr and none of the actions it considers can ever be carried out.

Three questions stay separate: is it allowed to work? What authority does it exercise? What are its conclusions worth? Disabling an agent says nothing about its quality, and promoting it does not make it any better.

Qualify before you trust

For every conclusion reviewed, you give a judgement: correct, incorrect (and which conclusion was expected) or impossible to judge with the evidence available. These judgements are the only evidence Kohentra accepts.

The thresholds depend on the action the agent proposes: the harder it is to undo, the more reviews are required and the fewer errors are tolerated. A judgement that has gone stale, because the subject has changed since, is excluded from the count and flagged.

A conclusion nobody has reviewed proves nothing.

A new version has to prove it

Improving an agent doesn't mean blindly replacing the old one.

  1. Preparea new version. A published version never changes.
  2. Evaluate side by sidethe same situations as the version in service, right after it, with no business effect.
  3. Compareon exactly the same input: agreements, disagreements, unjudged disagreements first.
  4. Check against your judgementswhere you had already decided. The cases that motivated the version are counted separately.
  5. Promoteonly on your decision. Nothing is inherited: no qualification, no permission.
  6. Keep measuringand roll back to an earlier version if needed, with the decision logged.

Autonomy is earned, one action at a time

Kohentra is read-only by default. For an action ever to be carried out, three independent locks must all open: the platform declares it executable, your workspace grants it to this agent in this version, and its regime allows it (propose, execute after confirmation, or execute without prior confirmation).

Today, exactly one action is executable: closing a ticket recognised as spam. Every other action can only be proposed. Executing without confirmation requires hard evidence, both on how reliable the triggering conclusion is and on how reliable the execution is. It stays supervised, logged, suspended at the first reported incident, and revocable at any time.

A security contract for every version

These rules are enforced by Kohentra while the agent works, never by an instruction given to the model.

  • What it may read

    the declared data and fields. Any other read is refused.

  • What may reach the model

    only part of what it reads, or nothing. Otherwise, no call is made.

  • What it may request

    its declared actions, reversible or not. Nothing else.

  • What you grant it

    the regime of each action, for this exact version.

  • What it will never do

    exceed the ceiling set by the platform, however reliable it is.

  • What it never sees

    no API key, no login, no token.

See the agents in real situations.

Invoicing, closure, tickets: what each one examines, and what it leaves for you to decide.

See the use cases Our security principles