Cognition’s October 19 update to Devin — introducing an Efficiency Score, Devin Coach, and auto-approvals tied to Agent Compute Units (ACUs) — marks a fundamental shift in how autonomous developer infrastructure manages enterprise cost.
When autonomous coding agents first entered enterprise stacks, organizations reacted with blunt financial instruments: uniform usage caps and monthly token budgets. This was a structural failure of governance framing. Treating agentic compute like a static electricity meter forced CFOs into budget panic while placing high-output engineers behind bureaucratic admin review queues.
The introduction of an "Efficiency Score" (calculating productive engineering hours per ACU rather than raw merged pull requests) redefines developer output. By giving credit for non-code labor—such as triage, system analysis, and architecture planning—the software moves the enterprise evaluation metric away from superficial volume and toward cognitive yield.
More critically, tools like Devin Coach install performative friction directly at the point of prompt execution. By flagging vague requests, over-powered model selection, or un-scoped tasks before a session runs, the system forces developers into a structured, plan-first habitus.
By replacing manual admin approvals with algorithmic auto-limits, Cognition has automated internal enterprise gatekeeping. High-efficiency developers are granted operational freedom, while wasteful prompt patterns are routed to human oversight. This transforms autonomous infrastructure from an un-auditable, volatile cost center into a self-policing institutional asset.