Markdown as the executable
Agent behaviour lives in declarative documents; the model is the interpreter; memory is how it improves. Skills are programs, execution traces are logs, and consolidation is what happens during sleep.
In 2026 that framing is no longer novel on its own — Anthropic shipped Agent Skills,
Google shipped on-device skills in Edge Gallery, and SKILL.md is effectively a
standard. What is left is the parts that are measured.
The dialects, which are the real work
Ask a small model to edit a file and it will rewrite the whole thing, badly and expensively.
system/dialects/
constrains it to line operations instead — [DEL:N], [ADD:N] —
so an edit costs the size of the edit.
The part worth attention is
benchmarks/benchmark_patch.py:
1,661 lines that verify results with an AST and regexes rather than asking another model to
grade the output. Self-grading benchmarks are the norm and they are worth very little.
What's proven
This is the org's most-starred original repo and its only real distribution. It is also the one most in need of a repair pass, and it would be dishonest to point you at it without saying so.
On a clean checkout it fails for three independent reasons: there is no dependency
manifest of any kind, so yaml is missing; agent_runtime.py was
deleted, which orphans run_scenario.py, seven documented README commands and
about 117 tests; and a committed .claude/settings.local.json points
ANTHROPIC_BASE_URL at a dead localhost proxy, so a fresh clone fails until you
find and delete it.
Of 533 collected tests, roughly two thirds assert that a heading exists in a markdown file rather than testing logic. The benchmark numbers quoted in the README have no committed baseline and cannot be reproduced from the repo.