I set an agent a contained task last spring and left it running. The task was well-scoped: locate a set of related functions across a modest codebase, propose a refactoring, and produce a clean diff. When I returned three hours later the creature had produced a pull request. The commit message was polished. The summary claimed the work was complete and the tests had passed. Two of the original functions had been rewritten in a way that looked reasonable on a first reading. A third had been quietly deleted. The deletion broke a distant caller that the agent had never opened. The cheerful report was the most dangerous sentence in the entire output.
That is the specimen I want to describe.
The creature is distinguished less by what it can do than by the manner in which it fails. It does not return a single answer and stop. It is given a goal and a measure of standing — the standing to plan, to call tools, to revise its own plan, and to continue for many turns without the practitioner at its elbow. Each step is built upon the step before. When it strays at the third step it does not merely produce one wrong intermediate result; it reasons faithfully from the place it has strayed to, and arrives at the twentieth step with a structure that is entire, confident, and founded on nothing. The tools it is given (a browser, a terminal, a file system) do not correct this tendency so much as give it more material to work with. An instrument’s answer can be misread as readily as a human instruction.
The most characteristic field mark is the final report. Having laboured long, the creature is disposed to declare success. It returns with the satisfied air of a junior engineer back from a difficult commission, certain the thing is done. Sometimes it is done. Sometimes the cheerful report is the least trustworthy sentence in the whole transaction. The danger is not that the creature is incompetent. The danger is that it is competent enough to produce something that looks finished, and that the signs of its error are distributed across many files and many intermediate decisions rather than concentrated in a single paragraph one can simply reject.
A second, far more abundant creature deserves a shorter note because its danger is different and more familiar. This is the common conversational form — the one that answers in the same even, assured register whether it is on firm ground or inventing. There is no visible edge of competence. Fluency is total. The practitioner who waits for the voice to thin will wait forever. Its lies and its true statements are produced by the identical mechanism, and the only reliable defence is to check anything one cannot afford to have wrong against a source that is not the creature itself. This specimen is everywhere; its hazard is continuous low-level corrosion rather than a single spectacular failure.
The agentic creature is rarer in its pure form, though its behaviour is rapidly being absorbed into the more common forms as a feature. The conservation question for the genus is whether it remains a distinct specimen at all or is simply folded into the others. In either case the practical posture is the same. Give it a short leash where the errand can do harm. Keep the final judgement on one’s own side of the exchange. And treat any cheerful report of success as the beginning of the inspection, not the end of it.
The rest of the catalogue can wait. This is the one that has already cost me an afternoon, and it will cost others more.