Dev uses an AI agent to audit its own code against Exo's talk — finds 3 concrete gaps
blixt · x · 2026-10-30
After watching Martin Casado's Exo talk, blixt had Coce (his meta-harness) audit its own codebase and surfaced three concrete design gaps:
- Toolboxes that outlive one chat: every automation run starts with an empty toolbox today; Exo's SREGym win came from one tool rewritten 5 times across incidents — sharing tool "branches" across runs makes automations cheaper over time
- A Guardian for an automation's changes to itself: runs can already edit their own prompt but nothing tests or rolls back; proposal — the next run trials the change, keep only if it completes without failure and cost doesn't rise, else auto-revert
- Let a run see its own cost: Exo cut its context after spotting cost in history; if runs could read earlier runs' cost, they could pick cheaper approaches automatically
Takeaway: these pay off only for repeated work (automations); tool-building in one-off chats adds overhead.
More from coding & agent
- 'People thought RAG would replace structured databases' — AI circle revisits LangChain hype — teodorio · 2026-10-30
- Tokyo Agent Hackathon: 40+ builders burn 580M tokens in 5 hours, ship 20+ projects — DavidBennett__ · 2026-10-30
- One-person AI music video: Suno + Hermes Agent + ComfyUI on a single RTX 3090 — Ok-Wolverine-5020 · 2026-10-30
- Many 'dumb agent' complaints are retrieval problems: RAG tool scores — rcleffke · 2026-10-30
- 1,656 commits in a month mocked as slop — CodeSandbox CEO jokes his own image is ruined — zeeg · 2026-10-30
- Self-Direction turns agent mistakes into LoRA lessons, lifting GAIA success from 38.8% to 55.8% — rohanpaul_ai · 2026-10-30