Summary
In this AI Engineer conference talk, Dotta — creator of Paperclip — challenges one of the most deceptively simple assumptions in agentic software development: that “done” is a binary state. When an agent opens a pull request, passes tests, and updates documentation, has it actually finished? Dotta argues that treating completion as a checkbox creates a dangerous failure mode in high-volume agentic workflows, where AI systems can generate more work than any human team can meaningfully verify — producing what he calls “verification theater” at scale.
Paperclip’s “liveness model” addresses this by decomposing task completion into a structured bundle of claims: artifact produced, evidence of completion, rubric met, reviewer assigned, authority to approve, and residual risk acknowledged. The talk introduces three core invariants for agentic control planes — productive work must continue, only real blockers should stop progress, and infinite loops must be bounded. Dotta walks through Paperclip’s specific mechanisms: first-class task blockers, human approval checkpoints with audit trails, and “watchdog” agents that act as goal enforcers across any coding harness — Claude Code, Codex, Hermes, or others.
The central takeaway is broadly applicable beyond Paperclip itself: stop modeling task completion as a Boolean and treat it as a structured object with distinct claims, evidence, scope, and ownership. For teams building or scaling multi-agent systems, this framework offers a concrete vocabulary for designing control planes that balance throughput with genuine quality assurance.
📺 Source: AI Engineer · Published July 12, 2026
🏷️ Format: Keynote Launch
