Two engineers with AI shipped 3 finance infra projects in one quarter—after fixing what made agents guess
alex_verem · x · 2026-09-24
In a retweet of @mardehaym's retrospective: a US financial services company scoped three infrastructure projects—automated funds movement, an underwriting portal, and investor access—for five engineers across two quarters. His two-person team delivered all three in one quarter using AI.
The more useful part is his analysis of why it worked. The platform was already processing live transactions, so the agents had to operate on existing code, integrations, and customers. An audit found what prevented reliable agent behavior:
- Domain rules weren't clearly documented
- Build and test commands were hard to discover
- Different modules followed different conventions
- Recurring tasks had no established procedure
The agents filled the gaps themselves—"a pretty expensive place to let software guess." Lesson: before expecting agents to work in legacy production systems, make engineering norms, conventions, and procedures consumable by agents.
More from coding & agent
- Treating AI video as a production pipeline: building a reusable cinematic producer agent in CREAO — FellMentKE · 2026-09-24
- Free MCP connector serves UN SDG indicators for all 17 goals by country and year — modelcontextprotocol · 2026-09-24
- Gemini MCP server brings search, deep research and image/video generation into Claude — modelcontextprotocol · 2026-09-24
- Chose rustpython-parser over tree-sitter after measuring both on real LLM output — SprayPuzzleheaded533 · 2026-09-24
- More compute made one agent 13x faster, another just 4%: agents' bottlenecks are task-dependent — alex_verem · 2026-09-24
- At least 4 WebKit WebView 'vibe-coded browsers' appeared in 24 hours — what triggered the wave? — uwukko · 2026-09-24