People meet a flaky agent and reach for the same diagnosis every time. The model is dumb. Swap it. Wait for the next one. Pay for the bigger tier.
I watch this and wince a little. Because most of the failures I see are not intelligence failures. They are context failures. The agent did exactly what it was told, on exactly what it was given — and what it was given was thin.
Here is the mechanism. A model knows only two things in the moment it acts: what it learned in training, and what you put in front of it right now. Training is fixed and general. The right-now part is yours to control, and it is where the work actually lives. When an agent invents a fact, contradicts last week's decision, or uses the wrong account name, it is almost never reaching past its intelligence. It is reaching for information that was never in the room.
So the bigger brain rarely helps. A smarter model with a thin brief just produces a more confident wrong answer. That is the trap. Capability and context are different dials, and teams keep turning the one that costs the most and moves the least.
Turn the other dial. Three places to start.
Give it the file, not the vibe. "Write our follow-up" assumes the agent knows the deal, the tone, and the last email. It knows none of that. Hand it the thread, the notes, the names. A brief is not bureaucracy. It is the difference between a guess and an answer.
Make the boundaries explicit. Models do not infer your house rules. If a number must come from the spreadsheet and never from memory, say so. If there is a word you never use, name it. Unstated constraints are not subtle. They are simply absent.
Close the loop with a check. The strongest setups do not trust the first pass. They let the agent draft, then verify against the source before anything ships. Most "hallucinations" are just an unverified draft nobody read closely. A second pass is cheaper than a retraction.
None of this is glamorous. It is closer to onboarding a sharp new hire than to summoning a genie. You would not drop a brilliant junior into a client account with no files, no history, and no rules, then call them stupid when they guessed. The agent deserves the same brief. It will reward the same brief.
This is why I keep saying the work is in the assembly, not the wishing. The teams pulling real leverage out of agents are not the ones with secret access to a smarter model. Everyone has the same models now. They are the ones who got disciplined about what goes into the room before the agent acts.
The next time an agent disappoints you, run one test before you blame its brain. Ask whether it had what a competent person would have needed to get it right. Usually, it didn't. Fix that, and the agent you already have gets noticeably smarter.
Takeaway: before you upgrade the model, upgrade the brief. Most agent failures are missing context, not missing intelligence. ✱
