Automation with AI agents
Controlled task execution by AI agents
In short: I built a system that lets AI agents carry out tasks in isolated environments. Permissions and scope come from a contract, results go through independent review, and accepted changes are passed to the target system in a controlled way.
01 / For your organisation
An AI agent can prepare an analysis, change code or propose an infrastructure change. To trust it with real work, you need clear permissions, independent review of the result and certainty that exactly what was accepted reaches the target system.
This way of working lets you use agents in processes where accountability for the result matters, and swap the model without changing the acceptance rules. A contract and evidence, not the executor’s own claim, decide whether work moves to the next stage.
02 / Delivered in 4digital’s environment
In 4digital’s own environment I started with requirements and the boundaries of responsibility before writing any code. The work rested on four principles:
- A contract instead of instructions — a task has a machine-readable contract: resources, permissions, cost limits, the required outcome and acceptance criteria. Ambiguity causes an explicit refusal, not a default interpretation.
- No evidence is not success — every state transition follows from a verified contract, permission and evidence.
- Autonomy within permissions — the executor chooses how to work, but does not change the goal, the policy or its own acceptance criteria.
- Separated roles — execution, evidence custody, result assessment, the operator’s decision and control of external effects are kept apart.
Every task goes through execution in an isolated environment, independent review and acceptance of the exact candidate. Review is done by a reviewer with a different identity from the executor, and the result is checked by an independent verifier. An accepted result reaches the target system through an intermediary that prevents double publication.
03 / Outcome
The same mechanism handles different kinds of work, from a code change to an analytical report, and executors from different model providers. Changing the model does not change the acceptance rules or how evidence is bound.
This is 4digital’s own demonstrator, tested in defined execution and delivery scenarios. Final delivery qualification used deterministic executors without model calls; trials with real models belong to earlier stages. This does not establish production readiness. I adapt the supervision mechanism to each organisation’s systems, permissions and requirements.
Value: the AI agent does real work, but a contract, evidence and independent review decide what is accepted and what reaches the target system.
04 / Selected evidence
- Five consecutive stages, each with a contract frozen before implementation and a separate acceptance decision.
- Deliberately wrong results fed in during every qualification were rejected.
- Executors from two model providers handled without changing the acceptance rules.
- Delivery confirmed by an independent read of the actual data from the target system; a repeated request does not create a second record.
- Source of experience
- 4digital’s own environment
- Discipline
- Agent systems
- Technology areas
- Agent systems · Permission control · Isolated execution environments · Independent verification · Idempotent delivery · C#/.NET · Python · SQLite · Podman rootless