Research paper 22

Switchboard Harness DSL: Runtime Lab Scenario Packets as Governed Agent Work Contracts

Agent harnesses are often described as prompts, tools, runners, or evaluation suites. For governed agent systems, that description is too loose. A harness must also say what work object exists, which operator may act, what evidence must be produced, which approvals or budgets constrain action, which proof closes the loop, and what packet remains afterward. This paper develops the Switchboard harness DSL as a local methods design around that problem. The central object is the Runtime Lab scenario packet: a declarative contract containing Exchange seed state, operator roster, Line fixtures, adapter posture, workflow expectations, evidence rails, approval and budget policy, proof commands, and archive format. The source evidence is design and implementation-adjacent rather than compiler-complete: the Runtime Lab 2.0 review names scenario packets as the reintegration target, the Switchboard harness benchmark contract separates closed-track proof from open-track comparison, and the eval harness types expose cases, triggers, hard checks, rubric checks, traces, reports, and comparison summaries. The contribution is a bounded language design for declaring governed agent work before execution. The claim ceiling is final-local methods design: a scenario packet can make agent evaluation inspectable and replayable without turning every scenario into a public benchmark or every successful run into release authority.