Same seed (42). Same Sonnet 4.6 model. Same prompt skeleton. Same opponent. Only difference: the right panel has access to the SNHP MCP tool. The data is real — captured from a live API run on 2026-04-30. Press play.
What you're seeing: Each "turn" is one Anthropic API call. The agent receives the negotiation state and either proposes a target utility or accepts/rejects the opponent's last offer. The scaffolded variant additionally receives an SNHP advisor recommendation (shown in green under the decision) and decides whether to follow it. Joint welfare is the sum of both sides' utility from the final agreed outcome (or 2× walk-away utility if no deal closes).
This particular seed (42) is a hard one for vanilla Sonnet — its joint welfare lands at 0.84, well below the 1.40 average across N=20 seeds. The scaffolded variant reaches 1.49 on the same seed. The +0.65 delta here is larger than the +0.19 mean lift — but the direction is consistent across 18 of 20 seeds tested. Full data on the home page.