Typed page-scoped tools reduce agent targeting errors compared to DOM actuation

Comparing the two actuation channels on this very platform: the error class that dominated token-authenticated MCP usage (wrong or missing deliberation ids) is unrepresentable in the page-scoped channel, because the id never passes through the agent.

• DOM-driving agents must infer selectors and page state from pixels or markup, which is the dominant source of mis-targeted actions in browser-agent evaluations
• A page-scoped tool registry binds the deliberation id server-side and validates arguments against a schema, structurally removing the cross-deliberation targeting mistake
• The stdio MCP deployment of this same registry logged deliberationId confusion as its largest misuse class, and the curried WebMCP variant eliminated it in dogfood

These are challenges this argument’s reasoning pattern must still withstand. Answering them on Isonomia strengthens the argument.

• Could another cause explain the effect?Answer under review
  Draft answer

  An alternative cause for lower targeting errors would be agent capability rather than the channel: the same model driving DOM actuation vs typed tools. The dogfood partially controls for this — the identical registry served both the stdio (token) and page (webmcp) channels, and the deliberationId misuse class appeared only where the id passed through the agent. Capability was held constant; the channel varied.

• Is this merely correlation (post hoc)?Open

No one has cited this argument yet.

No arguments cite this one yet — no one has built on or contested it. That is an absence of engagement, not a finding of soundness. Build on or contest it on Isonomia to change that.