https://ssss.company/ Learning AI agents in public: real code, real terminal output, honest limits.en Tue, 03 Nov 2026 12:00:00 +0000 https://ssss.company/article-19.htmlhttps://ssss.company/article-19.htmlTue, 03 Nov 2026 12:00:00 +0000Weeks of learning AI agents in public. Now we build the 12 homepage dials for real, one tested slice at a time, as our shot at a tech career. The plan, the first real test (14 passing), and what is still planned. https://ssss.company/article-18.htmlhttps://ssss.company/article-18.htmlMon, 02 Nov 2026 12:00:00 +0000Three SRE habits for AI agents: an error budget, every incident leaves a rule and a test, and a 15-minute trace check. The demo we built, 14 passing tests, and what is still planned. https://ssss.company/article-15.htmlhttps://ssss.company/article-15.htmlMon, 02 Nov 2026 12:00:00 +0000Agents read logs, docs and tool output, and any of it can look like an order. Label every input by source and check the whole plan. The plan check we built, 14 passing tests, and what is still planned. https://ssss.company/article-17.htmlhttps://ssss.company/article-17.htmlMon, 02 Nov 2026 12:00:00 +0000A prompt that says don't leak is a wish. Scan every outgoing send for keys and customer data, check the host, then send, redact or block. The leak-guard check we built, 16 passing tests, and what is still planned. https://ssss.company/article-12.htmlhttps://ssss.company/article-12.htmlMon, 02 Nov 2026 12:00:00 +0000A timeout only says the answer didn't come back in time. Read the record by run ID, then DONE, RETRY ONCE or PAUSED. The reconcile check we built, 12 passing tests, and what is still planned. https://ssss.company/article-11.htmlhttps://ssss.company/article-11.htmlSun, 01 Nov 2026 12:00:00 +0000A human yes should cover one exact, reviewed change. What an AI SRE design gets right, the approval-receipt check we built with 10 passing tests, and what is still planned. https://ssss.company/article-10.htmlhttps://ssss.company/article-10.htmlSun, 01 Nov 2026 12:00:00 +0000Two quick yeses approved two changes nobody meant. The rule we wrote, a small approval checker with 18 passing tests, the hole we found in our first version, and the approval gate our agents get next. https://ssss.company/article-09.htmlhttps://ssss.company/article-09.htmlSat, 31 Oct 2026 12:00:00 +0000An agent saying "saved" proves a message was sent, not that anything was stored. What we took from Anthropic's guide to agent evals, the read-back check we built, and the terminal output. https://ssss.company/article-08.htmlhttps://ssss.company/article-08.htmlFri, 30 Oct 2026 12:00:00 +0000A model's reasoning said it would not call any tools. Then it tried to upload a file. What we learned from Cohere's report, the plan-vs-trace check we built, and the terminal output that shows it working. https://ssss.company/article-07.htmlhttps://ssss.company/article-07.htmlWed, 28 Oct 2026 12:00:00 +0000A higher score can hide a broken safety case. What we learned from Langfuse Experiments, the publish gate we built, and the terminal output that shows it working.