SMASHED.
From a rough spec, ideas, and team descriptions to a fully functional buildout in less than an hour.
Pre-Clarity
building now
Request a live demo
Leave an email I can write back to. Or message me on X.
The previous build had 368 commits, eight green suites, and 218 passing tests, and died on the first keystroke.
The problem
Setting up multi-agent buildouts takes too long.
Nothing real starts until the setup is finished, and the setup is the work that should not exist.
The solution
Everything you already have to do with multi-agent systems is still here, with every variable forced to be exactly where you need it, exactly when you need it there.
| Section | Technical components |
|---|---|
| Specification intake and destination ownership | Directory, manifest, and Git-pinned input ingestion; typed specification proposals; structural validation; bounded revision; digest-named acceptance files; path harvesting and classification; forbidden-content checks; construction-usefulness checks; write-ownership fences; comparison of derived destinations against the specification. |
| Clarity and team derivation | Objective-bound empty seeds; sealed source specifications; role-registry and context schemas; five derived artifacts—roles, context vectors, spawn briefs, information needs, and Team DONE; fresh model exchanges; validation feedback; explicit proposal acceptance/renaming/refusal; immutable Confirm generations; selected-generation pointers. |
| Buildout identity and workspace isolation | Canonical checkout verification; resident origination handoff; atomic workspace scaffolding; buildout manifests; derived registries; separate mutable data and shared machinery roots; consistent team/seat naming; contained output paths; source fingerprints; destination coverage and occupancy checks. |
| Role context and package compilation | Role registry; typed deployment-context vectors; authored-value precedence; context hashing; research-backed package generation; immutable SQLite evidence reads with in-memory views; variable-policy resolution; package manifests; compiled grants; Confirm-bound charter material; explicit gaps where research evidence is missing. |
| Forced-frame coordination | Disk-derived object states; declarative action/precondition catalog; dependency-derived availability; seven views of one publication; executable argument schemas; semantic frame revisions; published-choice validation; stale-choice refusal; grouped actions; remaining-work selection; failure suppression and repair routing. |
| Evidence, provenance, and action transactions | Canonical JSON; SHA-256 identities; resident Ed25519 attestations; external authority journal; versioned action receipts; before/after snapshots; observed file effects; source-history segments; evidence-derived signatures; interrupted-action settlement; explicit postcondition outcomes; optional content-addressed storage. |
| Assembly, readiness, and launch admission | Confirm-selected artifact validation; per-seat package assembly; typed coordination graphs; coordination-envelope rendering; HOW/WHAT contract checks; Stamp; configuration drafting and adoption; inbox creation; charter transcription; Affirm; readiness analysis; Arm; sealed whole-buildout launch plans. |
| Resident runtime and process containment | Immutable installed runtime images; source/dependency identity checks; supervised Webapp, Write, and Derive services; fixed capability IPC; pinned Codex CLI invocation; isolated invocation configuration; subprocess groups; parent-death handling; timeouts; Linux bwrap filesystem boundaries; separate measurement containment; scoped cleanup. |
| Construction campaigns and agent communication | Sealed launch packets; Lead/Comms/subject-team topology; bootstrap messages; Markdown PROTOCOL inboxes; whole-inbox hash provenance; authenticated producer handoffs; dependency and write-conflict scheduling; per-seat process receipts; supervisor leases and heartbeats; launch-set reconciliation; one-shot execution boundaries. |
| Measurement and completion verification | Sealed Team DONE checking scripts; expected exit codes; bounded execution; artifact hashes; destination occupancy; producer/consumer joins; missing-owner and substitution checks; authenticated measurement receipts; cleanup qualification; checker inventories; mutation probes; verification-shard aggregation; GAPSMASH exploration and scoring. |
| Observation and operator interfaces | Read-only browser control room; validated last-publication snapshots; Derive activity reporting; launch-status board; allowlisted backend readers; contained artifact browsing; Confirm-generation views; explanation-coverage maps; hash-linked observer recordings; offline evidence validation. |
| Post-construction operations | Admission only after construction reaches CLEANED; reuse of sealed rosters; pending-mail planning; plan digests; serial campaign locks; fresh campaign identities; topology-bound message provenance; wake adapters; dispatcher/watchdog; freeze controls; standing-work receipts; separate standing measurement. |