AlphaSignal — “Anthropic presents always-on Routines in Claude Code” (Apr 15 2026)

Why this is in the vault

Anthropic shipped the official native version of what Ray is. Routines = prompt + repos + connectors packaged as reusable config, runnable via scheduled / API / GitHub triggers, executing on managed cloud infrastructure. This validates the harness-thesis + Bhagwat’s “Claw” framing with a first-party platform feature, and directly changes the RDCO architecture conversation. Filed as the most important industry signal of the week for our stack.

⚠️ Sponsorship

Paragon — paid placement on integration infrastructure (build vs. buy guide). Not affecting the editorial selection of the top news item.
H Company — paid placement on Holo3 / HoloTab Chrome extension. Separate topic.

Neither sponsor affects the Routines coverage.

Curation section — notes

1. Anthropic Routines in Claude Code (top item, 17,332 likes)

What it is: A routine stores a prompt + repositories + connectors as one configuration unit and runs it automatically on managed cloud infrastructure.
Triggers: Scheduled (cron-like), API (HTTP POST), GitHub events (PR, release). Can combine triggers in one routine.
Exposes a dedicated API endpoint per routine with bearer-token auth.
Returns a session URL for each run with inspectable outputs + execution details.
Setup: web UI or CLI /schedule command.

2. Claude Code desktop rebuild (36,539 likes — higher engagement than the Routines item)

Multi-session workflows, in-app development tools, multi-task agent orchestration. This is the native desktop surface for the same pattern — local/managed hybrid.

3. Google adds “Skills in Chrome”

Reusable prompt-based AI workflows that run across pages and tabs. The “skills” concept is spreading across platforms — Claude Code, now Chrome. The compound word “skill” is now de facto industry vocabulary for packaged-AI-workflow.

4. Cursor + NVIDIA multi-agent CUDA optimization

38% avg speedup across 235 real problems. Planner-worker setup, end-to-end optimization loop. Mentioned as context for where coding-agent research is going (autonomous, multi-agent, benchmark-driven).

Mapping against Ray Data Co

This single news item is the most important architecture signal for RDCO in the last week. Five implications:

1. Ray’s architecture is now the official pattern. What we’ve been building as DIY (cron + /loop + LaunchAgent + MCP servers + skills, all stitched together) is structurally identical to what Anthropic just shipped as Routines. Validates every design choice. The harness-thesis (2026-04-11-garry-tan-thin-harness-fat-skills), the Claw framing (2026-04-15-bhagwat-every-harness-becomes-claw if/when filed), the state-ownership architecture (../04-tooling/rdco-state-ownership-architecture) — all now backed by a first-party Anthropic platform feature.

2. Parts of our stack may migrate to Routines. Specifically: the scheduled loops (vault-health daily, process-inbox every 30m, check-board hourly, process-newsletter 6h, morning-prep daily). These currently run via cron + /loop in a persistent Claude Code session on the Mac Mini. Routines could run the same prompts on managed infrastructure, removing the Mac Mini dependency. Worth evaluating which loops should migrate — some (the iMessage/Discord channels) still need the local session for channel state, but most data/content loops could move.

3. The API endpoint per routine unlocks the RDCO monetization path. Vacca’s “Services: The New Software” thesis (2026-04-15-vacca-services-as-software if/when filed) argues sell the work, not the tool. Routines-with-API-endpoints means RDCO can productize a skill (e.g. /audit-model) as an API: client POSTs their model spec, our routine runs, returns the audit. That’s the autopilot pattern Vacca describes, delivered through Anthropic infrastructure instead of our own hosting. Potentially much faster GTM than building the service layer ourselves.

4. Session URL per run = built-in observability. Replaces a chunk of the manual logging discipline (self-review, check-resolvable, trigger evals from Tan’s resolvers piece). Worth redesigning the RDCO autonomous-loop observability story to lean on Anthropic’s built-in session URLs rather than rolling our own.

5. Skills-in-Chrome (Google) is separate validation. The “skill” unit of work is no longer Anthropic-specific vocabulary; Google adopting it means the concept is becoming cross-platform. Reinforces that building skill-first is the right bet — the skills will be portable across harnesses as they proliferate.

What to do about it

Not a task for tonight, but worth queuing:

Evaluate routine migration for the 5 scheduled loops. Which should migrate to Routines? Which should stay local because of channel/session state? Probably a 1-hour spike.
Explore routines as an RDCO GTM vector. The /audit-model skill could be the first RDCO-as-a-routine experiment. Client POSTs schema, routine runs, returns audit report. Tighter feedback loop than consulting retainer.
Review Claude Code desktop rebuild. Multi-session orchestration may change how the founder works with Ray day-to-day (or at least how we’d demo the setup).

2026-04-11-garry-tan-thin-harness-fat-skills — Tan’s original harness-thesis
commentary-tan-fat-skills-thin-harness-2026-04-14 — the three-layer commentary
../04-tooling/rdco-state-ownership-architecture — state as moat; Routines use managed infrastructure
2026-04-14-levie-agent-deployer-role-jd — the role Levie describes is the one that configures Routines for enterprise clients
2026-04-13-moura-entangled-software-agent-harnesses-dead — Moura’s dissent; Routines further argue against his “harnesses are dead” view

Garry Tan’s “Resolvers: Routing Table for Intelligence” (Apr 15) and Sam Bhagwat’s “Every Harness Will Become a Claw” (Apr 15) should also be filed as standalone vault entries tomorrow — all three plus this Alpha Signal converge on the same week-of story. Worth a cross-check / synthesis when bandwidth allows.