kernablekernable

For AI agent systems · EU AI Act

Every action of your AI agents, recorded and signed.

Agents act fast and on their own. kernable keeps the record: every action checked before it runs, stored as a signed event, and replayable when someone asks why.

[ The question ]

"What did the agent do, and who allowed it?"

Sooner or later a customer, an auditor or a regulator asks. kernable puts the answer in one place, the moment the agent acts:

  1. 01Logs scattered across prompts, tools and services→
    One ordered history of every action

    Each action an agent takes is a command; what it changed are events, in one log.

  2. 02An agent that can call whatever it reaches→
    Every action checked before it runs

    Your rules decide per request, and a missing rule refuses instead of allowing.

  3. 03"The model decided" when someone asks→
    Who, for whom, and on whose behalf

    Every action carries the agent's key, the person it acts for, and what caused it.

  4. 04Logs anyone with access could edit→
    Proof instead of promises

    Signed and hash-chained; kernable verify shows any change afterwards.

  5. 05Personal data copied into every log→
    Erasable when a person asks

    Personal fields go to an encrypted vault; an erasure removes them and leaves a proof.

[ How it works ]

Agents propose. kernable checks, records and delivers.

Your agents stay in the language and framework they are written in. They talk to the kernel like any other part of your application.

Agentany language
Python, TypeScript, …

Submits an action as a command, signed with its own key and naming the person it acts for.

Guardrailsmiddleware

Your rules check every action before it is accepted: allowed tools, limits, tenants. Fails closed.

Workeryour business logic

Carries out the action and returns the events: what changed, and why it was allowed.

Reactorsfollow-ups

React to events: start the next agent, ask a human, notify a system. Every step is recorded too.

  • Read after write. When an action returns, every view of it is up to date: no agent acts on stale data.
  • One action at a time. Two agents changing the same record run one after the other, never over each other.
  • What you record is your choice. kernable records the commands and events it handles; prompts and model outputs too, if you put them there.
[ Regulation ]

Record-keeping the regulations ask for.

kernable doesn't make an AI system compliant: that covers your organization, your risk management and your models. It gives you the record they all rely on.

  • Automatic logging

    Events recorded over the lifetime of the system, in order and tamper-evident.

    EU AI Act Art. 12
  • Traceability

    Who acted, for whom, under which rule, and with what result.

    ISO/IEC 42001 · ISO 27001
  • Automated decisions about people

    The full history behind a decision, for the human who reviews it.

    GDPR Art. 22
  • Right to erasure

    Personal data in a vault, erased by a signed event that leaves a proof.

    GDPR Art. 17

Each control, with where kernable stops: Compliance in the docs →

[ Consulting ]

Bring agents into production, with a record.

We help you design the actions, rules and history of your agent system, and build the first part on kernable with your team. Artur has built platform infrastructure for AI governance under the EU AI Act and ISO 42001.

Artur Susdorf
Artur Susdorf Author of kernable

Twenty years of shipping software, the last seven designing the architecture other engineers build on. Author of comby, the event sourcing and CQRS framework that carries multiple commercial platforms in regulated markets. kernable brings the same model to teams that work in any language.

LinkedIn →

[ Early access ]

Build on kernable before 1.0.

kernable is in private preview. Tell us what you're building, and we'll send you the binaries and the docs.