Building a JSON query language,end to end.
Rhizome takes a single specification and autonomously produces a non-trivial, correct program — a small expression language with a REPL — written, tested, and assembled by a field of agents with no central planner.
The problem
Most agent demos leave a thesis as words: that a multi-agent system can autonomously build non-trivial, correct software — and that distributed coordination reaches a quality a single agent can't. This run sets out to make that thesis observable.
The task: build rq, a small query and expression language over JSON. A lexer, a parser, an evaluator, a file mode and an interactive REPL. Numbers, strings, booleans, null, arrays and objects; path access like .users[0].name; operators with correct precedence; built-ins such as length, keys and contains.
An interpreter is the point precisely because correctness and edge cases are the substance — operator precedence, division by zero, out-of-range indexing, field access on non-objects, unicode, deep nesting, unterminated strings, coercion rules. Here “the verifier caught a real bug” becomes an observable event, not a narrative.
The finished report will carry its receipts: the public repository, the per-agent commit history, the run logs, and a green test suite — an open invitation to verify.
Specification
The target grammar (EBNF), the type system, path-access semantics, the built-in functions, and the coercion rules the run committed to up front.
lands with the runArtifact
The runnable program and how to reproduce it — build and test with a single command, plus REPL and file-mode (rq '<expr>' data.json) examples.
lands with the runSource & commits
A link to the public repository: the full commit history with author = agent id, the branch structure, and the timings of the run.
lands with the runVerification
1–3 documented episodes where the verifier mesh caught a real defect the implementer missed — before → found → after — alongside the final green test suite.
lands with the runMetrics
The run summary: number of agents, commits, and tests, wall-clock time, and an honest note on what was done without a human in the loop.
lands with the run