Working at the frontier: How Cognition trusts Claude Fable 5 to work through the night
02:00 · July 10, 2026 · Claude Blog
Summary
Cognition, the company behind the autonomous software-engineering agent Devin, has folded Claude Fable 5 into its core system. Devin is designed to tackle the kinds of engineering work that often remain on the backlog—large-scale codebase migrations, accumulated bug fixes, and features that repeatedly slip schedules. Because the generated code must run reliably in production environments used by both startups and Fortune 500 customers, Cognition evaluates new models primarily through extended internal testing rather than public benchmarks.
Earlier Claude releases had already raised the practical ceiling: Claude 3.6 Sonnet first enabled reliable multi-step tool use, tripling internal adoption. Yet a persistent constraint remained the length of time an agent could operate without drifting from its objectives. Previous models typically remained coherent for only minutes to an hour; beyond that point they would lose track of competing requirements, skim log files superficially, or assert the first plausible diagnosis. On complex migrations they sometimes completed the nominal task while introducing subtle downstream errors.
Claude Fable 5 removes much of that horizon limit. Engineers at Cognition report sessions that continue productively for as long as eight hours while the agent remains in the cloud, paging through internal debugging interfaces, articulating the invariants it intends to preserve, and explicitly noting what it does not yet know. On the company’s own Frontier Code benchmark—an “anti-slop” suite that penalizes code passing tests yet failing real-world integration—the hardest subset score rose from roughly 10 percent with the prior Opus model to about 30 percent, a gain that internal dogfooding confirmed rather than contradicted.
The longer, more stable context window makes previously planned capabilities feasible. Devin can now monitor a Slack channel, detect an issue without explicit mention, scan the relevant codebase, and surface a proposed fix. It can also watch production metrics and initiate triage on its own. Cognition views these proactive, cloud-resident sessions as the intended default for engineering teams and expects them to account for the majority of agent activity within a year or two.
Why it matters
This article is highly relevant for product teams and builders as it details the practical capabilities of Claude Fable 5 in agentic workflows. Dutch AI practitioners can leverage these insights to build more reliable, long-running autonomous agents for software engineering and incident triage.
