Agent Algebra playground · powered by the agent-algebra runtime
Run the algebra.
Type a formula and the playground will parse, canonicalize, lint, and attempt a conventional lowering to typed graph IR. Lowered programs can simulate offline; their model-backed atoms can call a real model. Lint, policy, and runtime checks still apply, and external tool variants need real host bindings beyond this teaching runner. Analysis is free and local. Running model calls costs tokens: bring your own API key.
Formula / agent file
ASCII works too: -> ~> -| |= -- => · strengths: ~ advisory, ! enforced · options: Lp{max=3} · variants: Tl[browse]
Provider
Key trust boundary: Run sends the key from the currently served JavaScript page to https://api.anthropic.com when Anthropic is selected, or to the exact OpenAI-compatible Base URL you enter. There is no application backend or proxy, but you are trusting this page, its repository, and the deployment account. Use a scoped, revocable, low-limit key created for this playground; keep storage off unless needed; clear it afterward. For higher trust, serve a reviewed local checkout. Analysis never needs a key. A strict connect-src allowlist cannot cover arbitrary custom endpoints, so this GitHub Pages deployment does not treat CSP as the key boundary.
Results
Hit ANALYZE to parse, canonicalize, and lint the formula — no key, no network.
The executable Graph IR appears here: nodes, bonds, loops, branches, state cells, scopes, and the plan digest.
The offline simulator walks the graph with synthetic values — watch gates block and verdicts traverse without spending a token.
RUN executes the plan through the durable runtime: per-node log, brokered model calls, effect receipts, exact prompts, and the full trace.