Delegation is not collaboration

Every mainstream agentic-coding harness is engineered to minimize friction-to-accept. The right design minimizes friction-to-reject, the friction to reshape, trim, and rewrite. Until the tooling inverts this, agents will keep producing spaghetti, and greenfield projects will suffer most, because greenfield is where every accepted diff becomes precedent.

Parent discussions

• Agent-Editor Protocol — the protocol this opinion is observation of

Connected concepts

• AI Pair Programming Workflow — the three-stage workflow (research, parse, implement) is what collaboration looks like when the tooling supports it
• Human Decision as Development Artifact — if the decisions are the artifact, a rubber-stamp UX produces code with no artifact at all
• The hidden tax of agentic coding is context switching — the incentive problem named here has a cognitive cost, measured there
• The smart-editor mode — diffs as a transient interaction mode — the proposed repair

The point

Cursor binds acceptance to Tab, the cheapest keystroke on the keyboard and the one your fingers hit reflexively. Aider commits each diff automatically; rejection requires a git operation. Claude Code renders the diff as terminal scroll with a confirmation prompt. In all three, acceptance is one keystroke; rejection is a sequence plus a context switch. The asymmetry is built in, and it is a design choice, not physics.

Harness vendors optimize for the metric they can see: accepted suggestions per hour. The result is a UX that casts the human as validator. A validator checks whether each diff is locally plausible. Nobody in the loop holds the global shape, because holding shape requires cheap rejection and reshaping, which is precisely what the interface taxes.

Greenfield makes this worst. In a mature codebase, existing structure pushes back on the agent. In a new project there is nothing to push back. Each rubber-stamped diff becomes the convention the next diff extends. Spaghetti is not an agent capability problem; it is the compound interest of frictionless acceptance.

The Agent-Editor Protocol in the vault is interesting here because its most important message is not edit_proposal but edit_accepted, which reports what the human did to the proposal, down to humanChanged: true and how many annotations were trimmed. Delegation asks “did you approve?” Collaboration asks “what did you change?” and feeds the answer back into the agent’s context. Whether the protocol actually delivers cheap rejection in practice is an open question, but the design intent is right: the diff the human produces is the signal. A harness that can’t see it isn’t collaborating; it’s dictating slowly.

Open questions

• Can friction-to-reject be lowered without raising total friction? Does inline reshaping scale past a few files, or does it collapse back into review?
• Is the “precedent effect” measurable? Do early accepted diffs predict long-run structural entropy in greenfield repos?
• What would an evaluation for collaborative (rather than delegative) harnesses measure, given that accepted-suggestion rate is precisely the wrong metric?