Speakers: James Peter (JustEvery)
Track: AI Engineering
Giving an agent a skill sounds straightforward: it loads a Markdown file, follows the instructions and completes the task.
Ours launched a long-running, paid process and told the agent to wait for the result. But agents repeatedly started the process, became sidetracked and rebuilt the output themselves. In one test, only 1 of 16 runs produced a fully correct result, while 12 of 16 paid for exports they never used.
The agents had gone rogue, and we thought we knew the fix. We rewrote the instructions. Then we rewrote them louder, again and again. The agents still got sidetracked.
The real problem turned out to be one line, and it wasn't in the prompt.
This talk retraces that investigation from the agent's side of the screen: why an agent's strangest decision can be its most reasonable one, and what agents need from the systems they depend on.
https://data.webdirections.org/ai-engineer-sydney/sessions.json#78f99404-213d-43cf-a126-12506ef62101
Full text on webdirections.org
.