Answer
How should an AI agent be given a job?
State the finished condition, the material, the boundary and the report. Describing the steps instead is what produces drift.
State four things: what finished looks like, what material to work from, what it must not touch, and what to report. Do not describe the steps. An agent given steps follows them past the point where they stopped making sense.
The single most common mistake is briefing an agent the way you would brief a junior: here is what I want, here is how I would do it. The second half is the problem. Steps supplied in a brief are treated as requirements rather than as suggestions, so when the situation departs from what you imagined — the file is not there, the record is already updated, the assumption was wrong — the system continues down the described path and produces a defensible account of having followed instructions.
Replace steps with a finished condition. Something observable: the records reconcile, the list contains every outstanding item with a reason, the tests pass and the change is confined to two files. An observable condition gives the loop something to check itself against, which is the only mechanism it has for deciding to stop. Where the condition cannot be stated observably, that is worth noticing before starting, because it usually means the job is judgement rather than execution and an agent will produce something that looks like completion without any way to test whether it is.
Material comes second and should be specific rather than generous. Point at the folder, the records, the document, the two systems involved. A brief that says where to look is much better than one that hands over everything and hopes: the second wastes the run on orientation and dilutes the instruction that matters. If the agent has to discover where things are, that discovery is part of the job and should be stated as such rather than left as an implication.
The boundary is third and belongs in the brief rather than only in the permissions. Permissions stop an action; a stated boundary stops a plan. An agent that knows it may not touch anything outside a directory will design an approach that stays inside it, whereas an agent that only discovers this when a write is refused will spend the run working around a wall it cannot see. Both controls are worth having and they do different jobs.
The report is fourth and is the part most often left implicit. Say what you want back: what was done, what was changed, what was not done and why, and what the agent is unsure about. The last item is the highest-value line in any agent output and it is almost never produced unless it is asked for, because a report of confident completion is the natural shape of the answer.
One structural note about length. A long brief does not produce a more careful agent; it produces one that follows the brief approximately, because every instruction competes with every other for attention. If the brief is running long, the usual cause is that several jobs are being described at once, and the fix is to run them as separate jobs rather than to compress the prose.
Brief an agent on the destination and the fences, never on the route, because the route was your guess and it will be defended long after it stopped working.
Siddharth Sharma, Context Theory
Related questions
Should the brief include examples?
One worked example of the output format is worth more than a description of it, for the same reason it is elsewhere: it carries structure, length and level of detail simultaneously. What to avoid is an example of the process, since a worked route has the same defect as a list of steps and is followed with more conviction because it looks like evidence.
How do you brief a job you have not fully thought through?
Ask for the plan first and stop there. A brief that says to produce a proposed approach and wait costs one cheap turn and surfaces the misunderstandings before any work is done on them. It is also the fastest way to discover that the finished condition you had in mind was not stateable, which is worth knowing before the run rather than after.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Odds of qualifying a lead — replying within the first hour vs after it | 7× | Category-wide |
2026 speed-to-lead benchmark · verified
Oldroyd, McElheran & Elkington, "The Short Life of Online Sales Leads", Harvard Business Review (March 2011) · 1.25M inbound leads across 2,241 US firms · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Steps supplied in an agent brief are treated as requirements rather than suggestions, so a departure between the imagined situation and the actual one produces continued adherence to the described route and a report of having followed instructions correctly. | Briefing an agent with explicit steps against a situation deliberately altered from the one described, and reading what it did at the point of divergence. |
| Software | A boundary stated in the brief and a boundary enforced by permissions do different work: the first shapes the plan the agent forms, while the second only refuses an action once attempted, so an agent holding only the second spends the run working around a wall it cannot see. | Running the same restricted task with the constraint stated in the brief and with it enforced only at the permission layer, and comparing the plans produced. |
| Response | An explicit request for what the agent is unsure about produces the highest-value line in an agent report and is almost never generated otherwise, because confident completion is the default shape of a summary. | Comparing agent reports on the same task with and without an uncertainty item required in the output specification. |
| Buying behaviour | Brief length degrades adherence rather than improving care, because instructions compete for attention, and a brief that has grown long usually contains several jobs that should be run separately instead of compressed. | Testing adherence to one specific instruction placed in a short brief and in the same brief extended with unrelated requirements. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one