Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

three places where generated work meets an old boundary

AI rewrites, skill optimisation and multi-agent handoffs all meet the same condition: an existing boundary still has to work.

Published Read 3 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-09-29

three places where generated work meets an old boundary

A C library still has callers waiting on the other side of its ABI. A skill file is being treated like a candidate that must survive a test. Two coding agents get put inside one harness and immediately inherit the usual question: who owns the trace when they disagree?

1. Google’s Rust rewrite still has to speak C

Google used Gemini in a pilot rewrite of giflib, the long-lived C library that processes GIFs. The target was modest by infrastructure standards: roughly 3,000 lines, no SIMD or assembly, a stable codebase. The requirement was not modest. The Rust replacement had to remain ABI-compatible with existing C callers.

That leaves the FFI boundary carrying the difficult bits. Rust does not make pointer lifetime disappear when the public interface is still C. Google describes dedicated building blocks for ownership at that edge, plus transparent validation and a rollback plan for service owners. The generated translation got the project moving; the work that earns deployment is proving that the old callers can keep living with it.

reading: Google Bug Hunters — “Scaling Memory Safety: AI-Assisted Rewrites of C/C++ Dependencies to Rust”

2. SkillOpt puts an agent instruction through a held-out gate

Microsoft’s SkillOpt starts from a slightly uncomfortable premise: an agent’s reusable natural-language instructions can be optimised without changing the model. It edits skills from trajectories, checks the candidates, and writes the survivor as best_skill.md.

Its “Sleep” companion makes the loop explicit: harvest past work, mine it, replay it, consolidate a proposed skill, then hold it behind validation. That is a better posture than treating a polished prompt as a finished object. A changed instruction is a behavioural change. It deserves the same suspicion as a changed function.

The repository does not settle whether these loops improve every agent. It does make the acceptance condition visible instead of quietly letting the agent rewrite its own manual.

reading: microsoft/SkillOpt

3. OpenRig makes the handoff a piece of software

OpenRig is a small public attempt to run Claude Code and Codex as one system. The headline is ordinary enough. Two coding agents have been passed between terminals, tabs and colleagues for a while. Putting the pair inside a shared harness is the actual move.

The moment that boundary exists, it has to hold more than a model choice. Which agent sees the other’s work? Who breaks a tie? What remains after a run, beyond two confident summaries pointing at different edits? The repository does not claim to solve those questions. It gives them a place to become engineering work rather than a handoff somebody remembers badly.

reading: mvschwarz/openrig

left on the table

SceneView has a broad 3D/AR SDK and the now-familiar assistant-facing files in its repository. It is worth opening if that is your stack. Today it is a catalogue entry, not a story with enough friction to carry a section.

voice-identity