A framework for everyone delegating work to AI agents: when to keep the loop tight, when to let it stretch, and what has to be true before you let go. By Robert Glaser.
Live at https://elastic-loop.robert-glaser.de
A small site built around one idea: the human-agent feedback loop has a size, and sizing it well is a skill. The hub introduces the framework; five pages carry the argument:
- Loops — tight, elastic, loose: three zones and how to size the loop for the task in front of you
- Why — the search-space shift that turns backpressure into a discipline
- Harness — the backpressure layers that hold agent output honest against the system and the product
- Grading — outcome grading as the new specification: tests, rubrics, scenarios, goldens
- Roles — what every role can check about agent output that nobody else can
A site that argues context has to be able to serve itself had better be agent-readable itself. So it is:
| Surface | Where |
|---|---|
| Markdown rendition of every page | /<slug>.md (the hub at /index.md) |
| Site index for LLMs | /llms.txt |
| All pages concatenated | /llms-full.txt |
| Sitemap | /sitemap-index.xml |
| Structured data | JSON-LD on every page, plus <link rel="alternate" type="text/markdown"> pointing to the markdown rendition |
npm install
npm run dev # http://localhost:4321
npm run build # static build to dist/Built with Astro, deployed to GitHub Pages on every push to main (see .github/workflows/deploy.yml).
Issues and PRs are welcome. The prose lives in src/content/pages/*.mdx, one file per page. Keep contributions in English for now. Small corrections (typos, broken links, factual slips) are easiest as a direct PR.
The Elastic Loop grew out of How Does Truffle Taste?, the canonical write-up of my talk that introduced the framework.
Code is licensed under Apache 2.0. Site content (the prose in src/content) is licensed under CC BY 4.0.