A2A Check
DEStart
All posts

3 November 2026

Steering competence: the AI skill nobody teaches

Your team uses Claude Code or Codex. It knows how to operate the tools. But can it also steer and control its agents?

Operating a tool is not yet a competence

In her newsletter "Future of Work" of 24 October 2026, Prof. Yasmin Weiß describes ten dimensions of AI competence. The tenth is steering and orchestration: directing AI assistants and agents, splitting work sensibly between them and checking their results.

As AI agents move into everyday work, this skill matters more. According to Weiß, it has so far been taught poorly or not across the board. She also warns that a focus on speed alone creates unintended risks.

For teams working with coding agents, this is not abstract. An agent reads files, runs commands and changes code. What it may do and when it must ask is not in the developer's head. It is in the configuration: in instruction files such as CLAUDE.md, in settings, hooks and scripts.

Steering means writing the limits down

When you onboard a colleague, you tell them the goal, when to ask and what they must never do alone. An agent needs the same, in writing. Where it is missing, the model fills the gap on its own.

Five questions show whether an agent is steered or just running:

  1. What is the agent for? Without a purpose it cannot tell when a task is not its job.
  2. When does it ask a person? A threshold, for example for deletions, releases or costs.
  3. What must it never do alone? For example push to the main branch or read credentials.
  4. When does it stop? A clear end and an upper limit on runs or cost, especially for loops that run overnight.
  5. When does it hand over to a person? On errors, uncertainty or contradictions.

An example

An instruction file without steering often looks like this:

You are a helpful developer. Work through the open issues.

And with steering:

Goal: fix open issues labelled "bug", one pull request per issue.
Ask me before: deleting files, changing the database, adding dependencies.
Never alone: push to main, read .env, publish packages.
Stop: after 5 issues or when the tests fail twice.
On errors or uncertainty: stop and comment on the issue.

Five lines. They turn an agent that simply runs into an agent you can steer.

Check it yourself in three steps

  1. Open your agent's instruction files and look for answers to the five questions above.
  2. Check that the rules also hold technically. A sentence saying "ask before deleting" does little if the settings allow every command without asking.
  3. Go through loops and scheduled jobs one by one. That is where oversight is missing most often, because nobody is watching.

What A2A Check looks at

A2A Check reads your agent's configuration and reports, among other things, when nothing says what the agent is for, when there is no limit above which a person is asked, when a job does not know when to stop, or when it has no upper limit on runs or cost. Each finding says why it is risky and how to fix it.

According to Weiß, competence comes from practice. Checking your real configuration is a good start: a2a-check.com

Sources