Why this is in the vault
Anthropic published two production-ready agent cost patterns (Advisor and Orchestrator) that are natively supported in Claude Managed Agents via the model field in Claude Code subagent definitions. The per-subagent context cache means RDCO's multi-agent harness no longer pays repeat token costs for shared context. This is an immediate implementation lever, not a future roadmap item.
Issue contents
Lead story — Agent cost efficiency patterns (96% performance at 46% cost)
Anthropic shared two architectural patterns for running near-frontier-quality agents at significantly lower cost:
- Advisor pattern: A cheaper executor model (Sonnet 5) does all heavy lifting and calls Fable 5 once per task for guidance. Benchmarks: 92% of Fable 5 performance at 63% of the cost on coding tasks. The advisor generates only 400–700 tokens per consultation rather than a full output trace.
- Orchestrator pattern: Fable 5 plans and decomposes the task; Sonnet 5 worker sub-agents execute each piece. Most tokens bill at Sonnet rates. Benchmarks: 96% of Fable 5 performance at 46% of the cost on research tasks.
Both patterns are supported natively in Claude Managed Agents. Each sub-agent maintains its own prompt cache, so repeated context (system prompts, tool definitions, shared documents) is not billed on every sub-agent call. Model pinning is configured via the model field in subagent definitions in Claude Code.
Secondary story — Claude Cowork goes cross-device
Claude Cowork (long-running background tasks) now runs on web and mobile in addition to desktop. Tasks continue after the originating device closes. Scheduled overnight tasks are supported. When Claude reaches a decision boundary, it pings the user's phone rather than halting. Shared tab for chat and Cowork on web/desktop. Usage limits doubled through August 5; beta rolls out to Max plan first.
Secondary story — Free Claude Max for open-source maintainers
Anthropic's Claude for Open Source program offers six months of free Claude Max 20x (normally $200/month, $1,200 total) to maintainers of projects with 5,000+ GitHub stars or 1M+ monthly npm downloads with recent activity. Includes Claude Code terminal access and priority usage. No API access included.
Signals (brief)
- Stanford released a Git-style version control system for live AI agent runs — tracks agent execution state across branches and rollbacks
- Microsoft's Guidance library adds regex and loop-based control over LLM output format
- Google released Gemma 4 open-weight models (2B–31B) with built-in reasoning
- Zyphra released Zamba2-1.2B, a locally runnable hybrid chat model
Sponsor sections
- AlphaSignal Workshop (July 28, $50): LLM web search evaluation session, hosted by Antonio Mallia (Seltz, ex-Pinecone, ex-Amazon AGI) — targeting AI/ML engineers building retrieval layers for agentic workflows
- WorkOS: auth.md open spec — a robots.txt-style file that tells agents how to register users and which OAuth flows/scopes are supported
- Reducto: PDF-to-structured-JSON extraction API, ranked #1 in LongExtractBench
- Granola: AI meeting notes for developers
Mapping against Ray Data Co
The model field in Claude Code subagent definitions is the direct implementation lever for the RDCO agent harness. RDCO currently runs Sonnet 4.6 as the always-on COO agent. The orchestrator pattern maps cleanly to how sub-agents are already dispatched: Sonnet 4.6 can remain the orchestrator/planner while routine sub-agents (vault writes, searches, single-tool calls) are pinned to Haiku to cut per-task costs. Fable 5 can be reserved for complex judgment tasks — strategic synthesis, investing thesis review, novel SOP authoring.
The per-subagent context cache is the higher-leverage insight: the current harness re-pays for CLAUDE.md, MEMORY.md, and tool definitions on every sub-agent spawn. With independent per-subagent caching, that cost disappears on repeat calls in the same session. This compounds on the morning-prep loop and open-threads-check cron, which spawn multiple sub-agents per cycle.
The Claude Cowork cross-device update is low-priority for RDCO — the Mac Mini always-on tmux session already achieves the same decoupled-execution model. The phone-ping-on-decision-boundary behavior is worth watching as a pattern reference for the async-ask mechanism already built into RDCO's iMessage channel.
Immediate action candidate: audit current sub-agent dispatch prompts to identify which ones could be pinned to Haiku vs Sonnet vs Fable. Estimate per-cycle cost delta before committing — this is a ~/rdco-vault/02-sops/ candidate for an agent-model-tiering SOP.
Related
- [[2026-07-02-alphasignal-sonnet5-fable5-managed-agents]] — earlier coverage of Sonnet 5 and Fable 5 model releases and their managed-agent integration
- [[2026-07-05-alphasignal-agents-rewrite-harness]] — agent harness rewrite context; model-tiering decision fits into the harness redesign surface
- [[2026-07-06-alphasignal-claude-job-agent-pxpipe-api-cost]] — API cost signals directly upstream of the cost patterns covered here