Factory AI's coding agents are outperforming OpenAI and Anthropic's tools on enterprise use cases, according to internal benchmarks shared by co-founder and CTO Eno Reyes. The difference comes down to context management and what Reyes calls "harness engineering."
Factory's "Droids" run model-agnostic, connecting to GitHub, Linear, Slack, and Sentry without forcing IDE or vendor lock-in. They handle sessions lasting eight to ten hours, managing context limits and tool calls across the full software development lifecycle. The team spent three years building the infrastructure to make this work at scale.
The pitch matters because autocomplete tools are already increasing code churn 20-40% as AI-generated code proliferates, studies suggest. Factory's approach differs: embed quality signals upfront. Their agents use linters, type checkers, security scanners, and complexity analyzers to validate work without human review. If the codebase lacks these signals, Droids can add them.
Reyes says most organizations have minimal validation signals implemented. Factory's agents are optimized for these signals, which Stanford research highlighted as the sole predictor of AI agent success. The company calls poorly structured codebases "slop code" and targets large, messy repositories where other tools fail.
The technical details matter here. Recent LLMs trained with reinforcement learning, including Anthropic's Sonnet 3.7, produce what developers call "code smells": CLI-biased output and architectural anti-patterns that require manual fixes. Factory claims their context compression beats both OpenAI and Anthropic's approaches, though they acknowledge enterprise needs differ from public benchmarks like SWE-Bench.
NEA and Sequoia back the startup, betting on "agent-ready" codebases as the next infrastructure layer. Reyes, previously at Hugging Face, frames this as industrial process rather than research: hundreds of small optimizations that add up to agents that actually ship.
The test: whether enterprises trust agents for production work. Early implementations remain research experiments for most organizations. Factory's wager is that quality signals baked into the process, not bolted on after, change that calculation.