Why do AI agents go off track on long, multi-step work?
Usually because one wrong assumption at the start is carried into every step after it. A long task multiplies whatever the agent believed at step one, so by the time anything looks wrong, a lot of work has been built on top of it. The cheapest fix is not more autonomy or more supervision. It is making sure the agent has the company's relevant rules before the first decision that matters.
One wrong premise, thirty steps of consequences
Two different things are being called failure. One is a step that breaks: a tool call that errors, a malformed argument, a loop. Those are visible and largely solved with retries and better tool definitions.
The other is a step that succeeds while being wrong, because the agent was working from a belief nobody checked. It picks a segment, an approval path, a naming convention, or a default that is not how your company does it, and then everything downstream is built correctly on top of that. Nothing errors. The work is simply not usable.
More autonomy does not create that mistake, it raises the price of it. The same wrong premise that costs you a paragraph in a chat costs you forty actions in an agent run.
Five fixes, all of which help
Each of these is worth doing. It is worth knowing which failure each one catches.
Smaller steps and a plan
Break the task down and check the plan first. It catches a lot. What it misses is a plan built on the wrong premise, which still reads as reasonable.
A critic or reviewer model
A second pass over the agent's output. A critic helps only if it has enough information to recognize the mistake. If the reviewer receives the same incomplete context as the worker, it may confidently approve the same bad assumption.
Evaluations
Generic evals measure general task quality. Company specific evals can test your own expectations when you encode them explicitly. Neither one supplies missing runtime context to an agent in the middle of a task.
Tighter prompts
Works, right up to the point where the prompt is a policy manual that has to be maintained by whoever remembers it exists.
A human check at every step
Reliable and expensive. If a person has to read every step, the delegation has been given back.
Where those are the right investment
If your agent is failing because tool calls error or the model loses the thread, then smaller steps, retries, and evaluations are exactly the work to do, and they will move the numbers. Plenty of agent problems are engineering problems and should be treated as such.
The reason to look elsewhere is a specific symptom: the run completed, every step looks sensible in isolation, and a person who knows the business reads the result and says that is not how we do this.
Three things supervision cannot recover
| The gap | Why more oversight does not close it |
|---|---|
| A review catches a wrong action, not a wrong premise | The action is consistent with the premise. To catch it you would have to already know the company's rule, which is the thing that was missing. |
| The rule exists, in someone's head | The person who would have caught it was not in the loop, and what they know was never written anywhere the agent could reach. |
| The cost lands late | By the time the output is reviewed, the work has compounded. Rework is not proportional to the size of the mistake. |
Resolve the relevant rules before the first consequential step
HexaHQ's job in an agent run is to supply the organizational context the decision depends on: the current rules, the corrections people have made, the conditions under which each applies, and the prior decisions on this kind of work. That happens before the step runs, not as a review afterwards.
It is deliberately not everything the company knows. Handing an agent a larger pile makes the same problem harder, because the rule that mattered is now competing with forty that do not apply. What arrives is the part scoped to the task and to the person the agent is acting for.
This is also why we describe AI as an extension of a particular person's work rather than a generic worker. An extension inherits that person's context and permissions, which is most of what makes their judgment good. A generic agent starts every task as a capable stranger.
An agent inherits its person's boundary
An agent acting for a person receives the context that person is meant to have. That matters for reliability as much as for confidentiality: an agent given the whole organization's rules has to guess which set it is operating under, and guessing is the failure this page is about.
The same renewal task, twice
First run: one wrong assumption, carried the whole way. Second run: the rule resolves first.
-
Run one, step one
The agent is asked to prepare renewal outreach for a book of accounts. It assumes every account can be offered the standard discount.
-
Run one, twenty steps later
Forty drafts exist. Every one is well written. The eleven multi-year accounts were never eligible for that discount, so a quarter of the work is thrown away.
-
Run two, before anything is drafted
The discount rule, the multi-year exception, and last quarter's decision on early renewals resolve with the task.
-
Run two, work that lands
The book is segmented correctly, the drafts are right the first time, and the renewal tasks are created in Salesforce under the owner's account once approved.
Questions people ask next
Why do AI agents make bad decisions?
Most often because a decision depended on something specific to your company and the agent did not have it. The model then does the reasonable general thing, which is the wrong particular thing.
Why do autonomous agents waste so much work?
Because the cost of a bad premise scales with the number of steps built on it. The waste is not the mistake, it is everything correctly derived from the mistake.
How do I keep AI agents from going down the wrong path?
Put the rules that constrain the path in front of the agent before the first consequential step, and keep the set small enough that they are actually applied. Checking afterwards is more expensive and catches less.
Does more supervision fix it?
It limits the damage and it removes the reason to delegate. A reviewer that receives the same incomplete context as the worker may confidently approve the same bad assumption.
Is this the same as an evaluation suite?
No, and you want both. Generic evals measure general task quality, and company specific evals can test your own expectations when you encode them explicitly. Resolved context tells the model what your organization expects while it is doing it.
Should we just give agents less autonomy?
That is one lever, and it trades away the value. The other lever is raising the quality of the context before the decision, which lets you keep the autonomy you wanted.
Related pages
Don't give the AI more context. Give it the right context.
Resolve the rules that matter before the decision that matters. Free to start.