Two currency watchers track the Anthropic platform and the wider agent ecosystem, write dated decision reports, and leave every adoption decision to you — so the framework never drifts a generation behind between manual catch-ups.
Watchers fetch only the delta since the last baseline. Adopting anything is a separate, owner-gated step — the baseline advances only after adopted work is committed.
model: assignment — mechanical work runs cheap, judgment work pins higher. Adapters map tiers per host./cli-watch tracks Claude Code / Anthropic releases; /repo-watch tracks a registry of external agent frameworks. Both write dated decision reports with self-attested egress receipts — one row per fetch.get_watcher_status MCP tool plus a step in /morning and /weekly surface days since the last run and undecided candidates.docs/capabilities.md records the verified Claude Code capability surface — models, effort, workflows, scheduling, hooks — re-verified every wave as the watcher's baseline.Scheduled watcher runs are report-only behind five independent layers — restricted tool profile, write fence, egress pin, fail-closed guard, and defang on ingestion — adversarially reviewed through nine rounds of bypass attempts. The guard runs every external step under a stall budget, so a hung guard denies instead of being skipped.
The fail-closed guard denies on any parse doubt or stall and writes only the run lock and today's report, fetched web content is defanged before it can reach a report, and validator checks 39–52 keep the whole path enforced.