The Prompt Is the Bottleneck: Why 70% of Agentic Sessions Will Start Without a Human in the Loop
At Cognition, the team building Devin has started tracking a ratio that most engineering organizations don't measure: how many of their agent sessions are kicked off by a human versus by a machine signal. Right now, it's 70% human, 30% automated trigger. Nader Dabit, a senior engineer at Cognition, argues that number is already inverting — and that the engineering prerequisites for that inversion are unglamorous infrastructure work that teams aren't prioritizing yet.
The core observation is almost too simple once you see it: most of the signals that trigger engineering work are already machine-readable. A test fails. An alert fires. A deploy spikes latency. A spec gets approved. Right now a human reads that signal, context-switches, and types a prompt that essentially restates what the signal already said. The engineer is acting as a relay between two systems that could talk directly. The prompt, in this framing, is the bottleneck — a manual translation layer that exists only because the agent doesn't have direct access to the upstream signal.
The engineering implication is concrete and, deliberately, unsexy: the prerequisite for reliable trigger-based autonomous agents isn't a better model or a smarter prompt. It's comprehensive unit tests (so agents can verify their own output), good inline documentation (so agents have context without asking), reproducible development environments (so agents can run code without custom setup per repo), and rich system context (so agents understand architecture and service interactions beyond a single file). That scaffolding is what separates an agent that opens a pull request from an agent that autonomously closes an incident at 3 AM — without a human in the loop at all.
Dabit's piece is grounded in first-party production data from one of the most active autonomous coding deployments publicly known, which gives the infrastructure checklist unusual weight. For teams asking what to build next before expanding agent autonomy, his answer is clear: the readiness condition isn't a capability problem, it's an infrastructure one.