⚠️ Sponsorship
Anthropic writing about how to build on Anthropic products. The thesis (fewer subagents, more skills and tool search) happens to route customers toward Claude skills, Managed Agents and Anthropic's own migration tooling, and the closing CTAs go to pricing and sales. The evidence is anecdotal: unnamed enterprise deployments, one "~2x faster" customer, and one support-agent regression seen in one of two runs. No benchmark dataset backs the "a single agent with skills often beat both alternatives" claim. Take the design heuristics seriously. The "it usually wins" claim is unproven.
Why this is in the vault
It is the vendor's current position on the exact design choice Ray makes all day: when to spawn a subagent versus keep work in the main loop. It also gives an eval-gated procedure for folding subagents back in, which Ray has never run on its own setup.
The mental model
Split by context, not by problem type. A subagent should exist because the work needs a fresh, discardable context window, parallelism, verification independence or separate permissions. It should not exist because an org chart or a task taxonomy has a box for it. Everything else belongs in one loop, with procedure carried by skills and hard limits enforced inside tools.
Summary (author's terms)
- Subagents remain a core pattern for verification, parallel work and modular permissions. Keep them when (1) a single loop measurably degrades from context bloat, (2) the work needs parallel, fresh, throwaway context, or (3) the agent can't choose among its tools or the system prompt contradicts itself.
- Across several enterprise comparisons, folding routers, planners and domain specialists into one loop with progressively disclosed skills and tool search often scored higher than both a single big prompt and a multi-agent design, usually at lower cost and latency. One customer saw about 2x speed. Three reasons are given: shared context beats summaries handed back by subagents; progressive disclosure keeps context lean (the article says Claude Code's system prompt shrank by over 80% for newer models); and models from Opus 4.5 onward can hold much more in one loop.
- Failure modes of multi-agent setups at large companies: context lost at handoffs, subagent sprawl with no owners, a "Frankenprompt" of conflicting rules, local evals that nobody owns end to end, and latency. Splitting by problem type causes a telephone game. Splitting by org chart is Conway's law at work.
- Design rules: put ordering and domain procedure in skills. Put hard limits such as validation and approval gates in tools, which return plain refusals the model can read. Group MCP tools by user intent. Tool choice degrades past roughly 30-50 tool definitions loaded at once, so defer the rest behind tool search.
- Migration, gated by evals: start with about 10 golden end-to-end examples plus per-skill evals. Next, call subagents as tools with a shared response schema. Then merge specialists that chain together, touch the same data or are usually called together. Finally, fold them in one at a time, check which individual tasks flipped rather than only the total score, and run each task more than once. Closing rule: "measure, change one thing at a time, and keep what holds up."
Mapping against Ray Data Co
Ray is already a single main loop that spawns subagents when needed (the same reading as [[2026-05-03-alphasignal-single-vs-multi-agent-systems]]), so most of this article confirms the current setup. The useful part is a test: for each subagent, does the split exist for context or independence, or only because the work has a separate label? Results by surface:
Where the split is justified (keep):
- Fresh-eyes critics (
sw-critic,verify-*,behavior-critic). "Verification" is the first use case the article lists for keeping a subagent. The critic's value is that it doesn't share the producer's context, so the article's main reason to fold (shared context beats summaries) works against folding here. [[2026-11-16-claude-code-team-verification]] makes the same argument from Anthropic's side. Strongest fit in the whole setup. - deep-research, one subagent per question. The questions are independent, run in parallel, and each pulls in large sources whose context is thrown away afterwards. This is case (2) as written.
- AGENTS.md hard rule 4 (over ~5KB → subagent returns only the extract). This is case (1), context bloat. It fits an always-on session that runs 24h+ with compaction, a cost that one-shot enterprise agents don't pay. [[2026-11-05-alphasignal-claude-code-subagent-cost-cut]] shows Claude Code applying the same discipline to its own internals.
- Workflow fan-outs and the per-article newsletter fan-out, on the context-isolation reading: the full bodies (5-100KB each) never enter the parent.
Where the article says fold back, or at least test folding:
- Hard rule 4 ignores what the extract loses. The article's main argument is that a summary handed back by a subagent loses things the main loop would have used. Hard rule 4 applies the same 5KB cutoff (about 1.2K tokens) whether Ray needs a fact from an artifact or has to reason over it: a founder's own draft, a spec, code he's about to change. For the second kind, the extract is the telephone game. A possible fix: scope the rule to "artifacts you only need an extract from" and allow direct reads when the artifact itself is the subject of the work. This is a hard rule, so changing it is the founder's decision. Ray has no measurement showing either version is better.
station-criticruns one subagent per critic axis. Every axis reads the same artifact and runs on every pass. That matches the article's merge criteria ("touch the same data", "usually called together"). The skill justifies the split as protection against one axis copying another's diagnosis, a claim nobody has measured. Recommended test, following the article's procedure: on the golden set, compare one critic subagent that receives all the axes (still fresh-eyes, still separate from the producer) against the per-axis fan-out, then check which verdicts flip, with at least 2 runs each. Keep the fan-out only if the merged critic scores worse.- Where the newsletter mapping gets written. The
process-newsletterskill says the mapping to RDCO is the value, and that judgment needs vault context the subagent doesn't have. Verified 2026-11-17: the skill dispatchesgeneral-purposesubagents (process-newsletter SKILL.md:33), not the cheapfanoutagent, so the model tier is fine. The open question is context: the subagent writes the mapping without the vault context the parent has. The article's pattern suggests a split: the subagent extracts and the parent maps, or the mapping is checked against the vault before filing. This needs testing before anyone changes it.
Brigade stations (spec → tests → code → critic). On its face this is splitting by problem type, which the article warns against. The spec → tests split exists for test independence ([[2026-05-12-rdco-pipeline-rlhf-shaped]], plan → tests → impl), and that is a verification reason, so it holds. The risk is the station-critic fan-out (point 2), not the four-station chain.
What Ray lacks to run any of this. The article's whole procedure depends on an end-to-end golden eval set with a single owner. Ray has per-note checks (the audit's I-checks, /self-review) and per-artifact critics, but no ~10-example end-to-end set that could show whether folding a subagent helped or hurt. Until that exists, every change above is a guess.
Already at parity: deferred MCP tools behind ToolSearch (the article's 30-50 definitions ceiling) and skills as progressive disclosure, with about 61 local skills loaded by description. Ray already follows the article's "procedure in skills, hard gates in tools/permissions" split (e.g. [[2026-11-13-building-always-on-agents]] write gates, auto-mode classifier).
Open questions
- Does a single all-axis critic match the per-axis station-critic's verdicts on a golden set? That is the cheapest fold experiment available.
- Should hard rule 4 be scoped by purpose (extract vs. reason-over), not just size?
- Who owns an end-to-end eval set for Ray, and what are its first 10 examples?
Related
- [[2026-11-16-claude-code-team-verification]]
- [[2026-11-05-alphasignal-claude-code-subagent-cost-cut]]
- [[2026-05-03-alphasignal-single-vs-multi-agent-systems]]
- [[2026-11-13-building-always-on-agents]]
- [[2026-04-15-thariq-claude-code-session-management-1m-context]]
- [[2026-06-04-agent-workflow-patterns-catalog]]
- [[2026-09-28-claude-dev-automating-eval-design-hillclimbing]]