I build agents that do the work, then write about what breaks.

Katalon True Platform's lead product manager on TesterArmy's open-source e2e AI testing framework: what the team got right, why enterprise QA in regulated industries like banking (BFSI) and healthcare needs independent validation of AI-generated tests, and how e2e can run as an execution lane under True Platform's versioned test cases, requirements traceability, and audit trail.

On 2 October Andrej Karpathy suggested asking an LLM to explain things in ASD-STE100, the controlled language aerospace uses for maintenance manuals. I added it to my agent harness the same afternoon, softened to 80%, together with an ELI5 opener and a format ladder that prefers a diagram or a page over prose. Then I measured my own Claude Code transcripts. Sentence length barely moved, because Opus 5.5 and my September anti-slop rules had already done that work. Passive voice and progressive verbs dropped, and the format ladder changed far more than the language rules did: the share of replies that came with an HTML page went up about eight times.

Between late August and late September I ran my day job through three Claude models in Claude Code: Fable 5.1, Opus 5, and Opus 5.5. I parsed every transcript: 114 lead sessions, about 8,600 model responses, plus the subagents they spawned. Opus 5.5 makes about 60% more tool calls per request than Fable, with close to half the tool error rate and 30% less time per request. Against the anti-slop rules I set in September, my gate would have refused 3.2% of its chat replies, against 30.5% for Opus 5 under the same rules.

Last Thursday I asked my agent for release notes and a demo video for our weekly release. Opus 5.5 built its own pipeline to do it: HTML scenes rendered frame by frame, sound effects tied to on-screen motion, music composed to the scene timeline, and real product recordings cut into explained demos. The feedback turned into seven rules. The pipeline is now the release-video skill in COG v3.14.0, next to browser-use's video-use, which does the other half of the job.