What Agents Actually Ship
The demos promise an autonomous coworker. The agents that survive a quarter in production are narrower, duller, and worth considerably more. Learn to tell the two apart before you build.
Every few months a video goes around: someone types one sentence and an agent books, buys, files, and reports. It is real, it works on camera, and it tells you almost nothing about what to build on Monday.
The useful question is not "can an agent do this once?" It is "can an agent do this two hundred times a week, unattended, at a cost and failure rate somebody will sign off on?"
Those two questions have very different answers, and the gap between them is where most agent projects die. A demo needs one good run. A production agent needs a distribution of runs you can live with, an audit trail you can read afterwards, and a bad day that costs less than the good days earn.
So before any of the engineering, you need a filter: which tasks are actually agent-shaped? This lesson gives you one, and it is unglamorous on purpose.
A task is worth an agent when all three of these are true. Two out of three is not enough, and each missing one points at a cheaper design.
- Multi-step: closing the task takes several actions, and the later actions consume what the earlier ones returned
- Live data: the right answer depends on the state of the world right now, not on anything fixed at build time
- Path unknown in advance: you cannot write the sequence ahead of time, because the sequence depends on what you find
And one gate on top, which is not about the task but about you:
- Checkable: after a run, somebody or something can tell whether it actually succeeded
Four jobs from the GH-North backlog are listed below. Which one is agent-shaped by all three criteria?
A. Water zone 3 for 10 minutes every morning at 06:00.
B. Convert 4,000 rows of sensor CSV into JSON for the analytics team.
C. Humidity alarms fired overnight in three zones. Work out which zones
have a real problem and which are faulty sensors, then open work
orders for the real ones.
D. What is the current moisture in zone 5?- AC: several actions, each one chosen based on what the previous reading showed, and no sequence you could write in advance
- BA: it runs every day and touches real hardware, which is exactly the kind of repetitive work agents exist for
- CB: 4,000 rows is far too much for a person, so it needs autonomous handling
- DD: it needs live data from the greenhouse, which is the criterion that matters most
The mirror image is just as useful. Four task shapes come up constantly and none of them wants an agent:
The single lookup. One call, one answer. Wrapping it in a loop buys you latency and a chance to hallucinate the zone number. The fixed recipe. You already know the steps and their order. Every extra decision you hand to a model here is a decision that can now go wrong in a way your tests cannot enumerate. The bulk transform. Four thousand rows through the same rule. Volume is an argument for a script, not against one, and a script gives you the same output twice. The unverifiable judgment. "Decide which zones are performing well." If nobody can say afterwards whether the answer was right, you cannot operate it, tune it, or defend it. Agents are not a way to avoid defining success.A fifth non-shape is worth naming because it is seductive: the task that is genuinely agent-shaped but whose failure is unrecoverable and silent. Agent-shaped is a necessary condition, not a permission slip.
Here is a requirement as written by the operations lead. Assume it is implemented exactly as specified. Which of the three criteria does this task fail?
Every night at 23:00:
1. Read the moisture value for zones 1 through 6.
2. Write all six values to the metrics table.
3. If any zone is below 25 percent, email the grower a summary
listing those zones.- APath unknown in advance: every step and its order is written down, including the one branch
- BMulti-step: it is really a single action repeated six times, so there is only one step
- CLive data: metrics tables are historical, so the task does not depend on current state
- DIt fails none of them; the branch in step 3 makes the path genuinely unpredictable
So what does actually ship? Four shapes, none of which make a good demo video:
Triage and dispatch. Something noisy arrives. The agent gathers enough context to classify it and opens the right ticket with the right details attached. Humans still do the fixing. The value is in the twenty minutes of lookup, not the decision. Reconciliation. Two systems disagree. Investigating each discrepancy takes a different number of lookups every single time, which is precisely why nobody ever managed to script it. Bounded operation with a gate. The agent does the ten reads, the diagnosis, and the draft, then stops and proposes one irreversible action for a human to approve. Almost every agent that survives contact with a real operations team has this shape. Long-tail automation. A job too varied to script and too small to staff. Individually trivial, collectively a person-week a month.The common thread is uncomfortable if you came for the demo: narrow scope, mostly reads, and at most one gated write. That is what a quarter of production does to an agent design, and you may as well start there.
Your team proposes an agent that, on any humidity alarm, investigates the zone and then autonomously adjusts vents, irrigation, and the CO2 valve to correct it, with no human involved. Which of the four production shapes should you steer this toward first, and why?
- ABounded operation with a gate: let it do the investigation and propose the corrective action, with a human approving anything irreversible
- BLong-tail automation: alarms are varied and low-volume, so full autonomy is the cheapest path
- CBulk transform: the readings can be processed in batches, which removes the need for a gate
- DLeave it fully autonomous but add a longer system prompt telling it to be careful with the CO2 valve
- A task is agent-shaped only if it is multi-step, needs live data, and its path cannot be written in advance. Two out of three points at a cheaper design.
- On top of the three, one gate: the result must be checkable, or you cannot operate what you built.
- Four shapes that are not agent work: the single lookup, the fixed recipe, the bulk transform, and the unverifiable judgment.
- What actually ships: triage, reconciliation, bounded operation with a human gate, and long-tail automation. Narrow, read-heavy, one gated write.
- A branch you can enumerate at build time is an
if, not a reason to add a model.
Next: the six parts every agent has, and why every agent bug you will ever chase lives in exactly one of them.