Why this is in the vault
Reporting from inside Every's own team on how they've started informally organizing Anthropic's model lineup into a hierarchy (Fable as CEO, Opus 5 as senior engineer, Sonnet 5 as junior engineer/analyst), and on the "Claudish" dialect problem that emerges when agents write for other agents instead of for humans.
Mapping against Ray Data Co
Direct validation of the model/effort pairing already codified in feedback_delegation_model_effort_pairing.md: this session runs on Sonnet 5 at medium effort while Fable delegations get high/xhigh reserved for meaningfully large tasks — Every engineer Marcus Moretti's "Fable is the CEO, Opus 5 is a senior engineer, Sonnet 5 is a junior engineer/analyst" framing is the same mental model, arrived at independently. Head of platform Willie Williams's line — "we're moving toward more of a mixture-of-models structure" because "the qualities that make a good orchestrator don't make a good executor" — is the argument for why RDCO already routes coordination to one tier and execution to another rather than using a single model everywhere.
The sharper hit is the "Claudish" problem. Cora GM Kieran Klaassen's theory that Opus 5 was trained mostly to talk to Fable, not humans, so "it doesn't matter that Opus has linguistic patterns that humans find abrasive" — and head of consulting Natalia Quintero's complaint that subagent reports read "like gibberish," fixed only by building skills that "translate agent output back into English" — is exactly the failure mode CLAUDE.md hard rule #4 (route long artifacts through subagents, return only the summary) and feedback_workflow_agent_output_integrity.md (7 observed factory failure modes, including pointer-not-content returns and angle-bracket StructuredOutput corruption) already exist to prevent. Naveen Naidu's fix — forcing the agent to explain itself with "I don't understand the issue here, can you help me understand it?" — is a cheap, generalizable prompt worth keeping for any Ray subagent report that reads as unparseable shorthand instead of a real report.
The core argument
Frontier labs are moving from marketing a single generalist model to running a mixture-of-models system where each tier specializes — a coordinator model, a strong-but-abrasive senior-engineer model, and cheaper junior models for routine work. That specialization is efficient for the system but produces a side effect: models optimized to talk to each other develop compressed, jargon-heavy "Claudish" that is unreadable to the humans who eventually need the output. Every's own team is already building workarounds (explicit "explain it to me" prompts, dedicated translation skills) rather than treating it as a solved problem.
⚠️ Sponsorship
The email is majority sponsored/promotional content: an ElevenLabs ad block (ElevenAgents voice/chat platform, framed as "brought to you by" with a "want to sponsor Every?" footer link) plus Every's own Builder Pack self-promo (Paper Pro + Mobbin credits bundled into the $625/year All Access tier). The reported content itself (Fable-as-CEO framing, Claudish dialect) is not sponsor-influenced, but this was also only the free-preview portion of a paywalled post — the full piece (which models Every's team is using this week, a company-context-packet prompt) sits behind the subscription paywall and wasn't fetched.
Related
- [[2026-06-15-indy-dev-dan-claude-fable-5-banned]]
- [[2026-07-22-every-context-window-ai-team-launch]]
- [[2026-05-15-every-team-agents-vs-personal-pets]]