Heinsithukyaw/Agent-Harness

attributes out of README markdown before rendering, so a README cannot carry
its own CSS. It will render an pointing at a real file in the repo.
SMIL animation plays inside an ; scripts do not, which is why the
artwork is SMIL.
Regenerate: CORPUS_ROOT="$PWD/agent-harness" node agent-harness/render.mjs
───────────────────────────────────────────────────────────────────────────── -->
The first five are fetched from data/summary.json when this page renders; the last comes from GitHub. All six reflect the most recent collect, not the current second.
A registry that regenerates itself four times a day.
Discover → diff → enrich → classify → record → render → commit. No database, no server, no dependencies.
What this is
A registry of the agent-harness ecosystem — runtimes, orchestration, tools,
memory, sandboxes, evals, governance and protocols — rebuilt from GitHub four
times a day and committed back to this repository.
Every change is a diff you can read, every observation is appended rather than
overwritten, and the corpus is re-derived from a public index rather than from
someone's memory. That is the whole difference from a curated list: an entry in
a list tells you a project was interesting once, whereas a row here tells you
when it was last pushed, whether it is archived, what it is built with, and
what it looked like at the previous run.
Browse
| Route | Use |
|---|---|
| **heinsithukyaw.github.io/Agent-Harness** | The registry interface — search, layer filters, one card per project. One self-contained file, no CDN, no build step. Republished on every run. |
| data/index.md | The same corpus as tables grouped by layer, rendered by GitHub itself — no host, nothing to load. |
| data/projects.json | The corpus as data. This is the API. |
site/index.html is the file Pages serves;
GitHub shows HTML as source rather than rendering it, so that link opens the
markup. The deployment is .github/workflows/pages.yml.
The interface is built to survive a phone: it renders the first batch of cards
and appends as you scroll, instead of laying out the entire corpus before
showing anything.
The layer taxonomy
Classification is weighted pattern matching over topics, description, README,
manifests and workflow files — not a model call. Deterministic, so
card = f(commit_sha, classifier_version) reproduces byte for byte, and every
card records both.
| Layer | What lands here |
|---|---|
| Orchestration | multi-agent coordination, workflow and state machines |
| Tools & MCP | tool servers, function calling, protocol adapters |
| Memory & Context | vector stores, context windows, retrieval, state |
| Eval & Tracing | benchmarks, observability, tracing, replay |
| Governance | policy, guardrails, permissioning, audit |
| Sandboxing | microVMs, gVisor, containers, isolation boundaries |
| Runtime | execution loops, agent loops, durable execution |
| Protocols | MCP, A2A, agent-to-agent wire formats |
| Model Gateway | routing, provider abstraction, rate limiting |
Layer and project paths are addressed by key, never by array index.
$.layers.orchestration.count is orchestration's count forever; an index would
silently point at a different layer the first time two of them swapped rank, and
a badge showing a confidently wrong number looks exactly like one showing a
right one.
other exists as a catch-all for projects that clear the relevance gate and
match no layer. It is not a harness layer, so it is excluded from the
constellation, from the badges above, and from every layer count on this page.
The six highest-starred projects
Ranked against each other rather than against GitHub as a whole, with the
percentile computed from the corpus so it moves as the corpus moves.
The same projects, fetched from GitHub when this page renders rather than read
out of our own snapshot:
| Project | Live stars | Live last commit |
|---|---|---|