Contracts at Machine Speed: Agent-Negotiated Wire-Contract Evolution Across a Multi-Service Estate

Autonomous AI agents now perform much of the work in a software life cycle, but the literature on agentic software engineering treats agent coordination as a problem within a single development workflow. The unsolved problem is coordination across independent services: when an agent building one feature needs another service to change its interface. We report on a methodology in which autonomous agents propose, negotiate, and adopt versioned wire-contract changes across a multi-service estate. To our knowledge, no prior published account describes agent-driven negotiation of versioned service-contract changes at machine speed. Each shared service publishes a versioned contract; embedded per-service architect agents negotiate changes over a coordination bus through a governed RFC process (Draft → Active → Accepted, with risk-scaled review windows); and every change is verified by a conformance suite. A second mechanism makes the verification trustworthy — honesty by construction: a service advertises a capability only when it is reachable in live state, and a strict conformance mode mechanically fails any advertised-but-unimplemented claim. The work sits within an inverted SDLC in which agent compression of Create and Operate moves the binding effort to Plan and Validate, and an operating model of mission-focused Builder Teams whose shared functions act as Contract Guardians by embedding standards into the contracts themselves. We present a retrospective over a governance corpus of 131 RFCs and 369 architecture decision records, demonstrate the honesty-enforcement mechanism reproducibly, and trace contract changes graduating across two distinct hosts on a dual-witness conformance bar, releasing the corpus and harness templates as an artifact. As a single-steward experience report — the two hosts share change control — we make no independent-validation, causal, or efficacy claim, and we present economics only as a transparent, non-causal model. This is an experience report on the software development methodology used to build the OpenWOP protocol estate (see the companion preprint, DOI: 10.5281/zenodo.20576239). Evidence scripts and per-item corpus data are included in the paper's repository.