[smithclay]

Prep for the Build the Next-Gen Git Platform competition (closes Oct 14, 11:59 PM PDT).

Thesis

Code memory for Code Mode agents, on Cloudflare Artifacts. Throw away the isolate, keep the code. The video lands this as a one-two punch:

1. Code Mode: models write code better than they call tools, and that code is "a compact plan" (Cloudflare's words). Each plan runs in a disposable isolate, so it's thrown away after the run.
2. Code memory:
   • Phase 1: every execution is saved to the agent's own Artifacts repo.
   • Phase 2: Workers AI embeddings and Vectorize let the next agent start from the closest code others already ran.

The script opens with an alice→bob walkthrough on the example's catalog, then uses DABstep as the deep dive.

Changes

• DEMO.md: an 8-minute run of show built around the thesis. It opens on the learning curve, then covers durable Code Mode, memory at fleet scale (the swarm), learning from memory, which kind of memory works (raw runs vs distilled skills), results, and use plus what's next. Each beat maps to a judging criterion and has a draft voiceover, plus an asset checklist and pre-flight list.
• Learning curve: benchmarks/dabstep/scripts/learning_curve.py and a new section in docs/experiments.md, computed from Adyen's official per-task scores.
  • Tasks 1–8, with no memory yet: both arms use about 315–320k median tokens.
  • Tasks 25–48: 45k vs 215k.
  • Tasks 97–450: 60k vs 186k.
  • Accuracy is 11–19 points higher in every fifth of the run.
• docs/submission.md: draft answers for every field on the submission form, with the vision rewritten around the thesis.
• examples/agent/scripts/swarm.mjs: the concurrency burst. It starts N agents at once (one Durable Object, one repo, one commit each) and reports commits/s and p50/p95 latency.
• Example cleanup: drops the snippets endpoint and updates the catalog comment.
• README: leads with the thesis and documents the swarm script.

Verification

• npm run typecheck and npm test pass.
• The swarm script ran against a local mock server: 100 agents, with injected failures counted as errors.
• It has not run against Artifacts yet. That's the next step, before Oct 15 when Artifacts billing starts. Its numbers fill in the bracketed placeholders in the script.