Your AI agents don't have a goal problem. Your organization does.

Every team deploying AI agents in 2026 hits the same wall, usually within weeks. The agent is capable. The model is state of the art. And the results are mediocre — meandering work, confident nonsense, loops that never finish or finish wrong.

The engineering diagnosis is nearly always the same: the goal was badly specified. Too vague to act on, no definition of “done,” no way to verify success. The fledgling discipline of loop engineering is blunt about it: an agent needs a goal specific enough to act on, real feedback from its environment, and a testable termination condition. Without those, the most capable model in the world just generates expensive motion.

Now here’s the observation I can’t unsee after 26 years inside organizations: that is a precise description of how organizations fail their people.

The mirror

Take the goals your organization actually issues. “Improve customer experience.” “Become data-driven.” “Adopt AI across the business.” “Increase operational excellence.”

Try to hand any of those to an AI agent. It would fail instantly — and now we can say exactly why it fails: no actionable specificity, no context, no verifiable outcome. The machine exposes in seconds what humans absorb quietly for years. Because that’s the thing about people: they don’t return an error message. They nod, interpret the fog as best they can, drift in divergent directions, and generate the organizational equivalent of a bad agent run — busy, plausible, unverifiable. We call it “execution problems.” It’s specification problems, wearing a disguise.

I’ve watched this through five technology waves, and the pattern holds: teams almost never lack talent. They lack a goal that would survive being given to a machine.

The best goal I ever worked toward

The clearest goal of my career was at the European Parliament. Not “improve the website” — but: raise the task success rate of the navigation, measured by real user tests. When we started, that rate was under 20%. Fifty-plus stakeholder workshops, thousands of user tests and many design iterations later, it was above 82%.

Notice the structure: specific enough to act on, feedback from reality (6,000+ user tests — not opinions), and a testable finish line. In 2026 terms, that goal was agent-runnable. It’s not a coincidence that it succeeded at continental scale, across 27 languages and peak traffic of a million visitors a day, in one of the most complex institutional environments in Europe. Well-specified goals survive complexity. Vague ones don’t survive contact with a hallway.

The exercise: write the agent brief

Here’s a practice I now push leadership teams toward, and it works even if you never deploy a single agent.

Take your most important initiative and write it as a brief for an autonomous agent, with four sections:

1. Goal — what outcome, stated so an entity with no shared history and no ability to read your mind could act on it.
2. Context — everything the agent needs to know: constraints, history, dependencies, definitions. (Can you even assemble this? If not, you’ve found a deeper problem.)
3. Tools & permissions — what it may use and decide alone, and what requires escalation.
4. Done condition — the test that verifies success, run by something other than the agent’s own judgment.

Most executive teams cannot complete this for their flagship initiative. The goal dissolves under specificity; the context turns out to live in six heads; nobody agrees on the done condition. That discovery stings — and it’s worth a year of steering committees, because it shows precisely where the fog is.

Then the reframe: your people have been operating without this brief all along. They compensated with intuition, goodwill and corridor conversations. The agents just refuse to compensate — which makes them, unexpectedly, the most honest audit of organizational clarity ever deployed.

The machines are holding up a mirror. The organizations that look into it will get something better than automation: they’ll finally get clear.

Sources

• Tosea, Loop Engineering Guide 2026
• Sourcegraph, Context Engineering