Cognition launches Devin Fusion hybrid-model harness
Cognition has introduced Devin Fusion, a hybrid-model harness that dynamically routes tasks mid-session between a sidekick agent and a frontier model. This approach reduces the cost of frontier-level intelligence by 35% while maintaining performance on complex coding tasks.
Devin Fusion tackles a major pain point in agentic AI: the high cost of frontier models and the inadequacy of smaller models for complex, ambiguous tasks.
- –The dual-agent "sidekick" architecture is a smart evolution from basic LLM routing, allowing for continuous delegation and review.
- –Dynamic mid-session routing is crucial for software engineering, where task complexity is often hidden until the work begins.
- –Focusing evaluation on PR merges (FrontierCode) rather than traditional benchmarks shows a commitment to real-world utility over theoretical performance.
- –This update significantly improves the unit economics of AI coding agents without sacrificing capability.
DISCOVERED
2h ago
2026-06-29
PUBLISHED
2h ago
2026-06-29
RELEVANCE
AUTHOR
cognition