Sentry CEO on agent determinism: why system prompts are almost never the answer
zeeg · x · 2026-10-30
Sentry CEO David Zeeg's essay 'A Little Determinism' tackles a core agent-engineering problem: creating determinism where none exists — getting agents to behave as you would without micromanaging. Key points: adding constraints often degrades agent capability; skills/AGENTS.md/system prompts are almost never the answer; he details how Sentry's Slack-driven agent 'Junior' evolved from turning threads into GitHub issues to how he ships software, with a sandbox-to-PR workflow. Models carry the goal once given the right tools (currently Opus 5.5 is migrating Sentry to Rust).
Related event: Sentry Founder's Essay: Bringing a Little Determinism to AI Agents(2 posts)→
More from coding & agent
- A growing collection of CSS view transitions with copy-for-agent snippets — jh3yy · 2026-10-30
- GitHub Copilot with Opus 5.5 beats Claude Code with the same model, dev claims — DimitrisPapail · 2026-10-30
- Two agent runs crashed on their own plan files—7 self-repair rules to save the whole run — petrusenko_max · 2026-10-30
- NVIDIA releases open Nemotron 3.5 Embed, Rerank and Parse models to cut agent token costs — NVIDIAAI · 2026-10-30
- The Infinite Work Machine: buyer bots vs seller bots turn AI efficiency gains into more work — Ididit-forthecookie · 2026-10-30
- Seroter's reading list: agent identity management, AI code review fails, skills as products — rseroter · 2026-10-30