Agents need runtimes, audit trails, and workflow tests

Autonomy Kernel gives the strongest design statement of the day. It defines a runtime under agents and models that owns execution, identity, authority, communication, and auditing. Every action must trace back to one principal and one authorization path before it runs.

The proposal is useful because it names the missing operating layer for long-running agents: scoped permissions, leases, audit records, portable state, and a guaranteed stop path. The source does not report an implementation or benchmark, so its value is architectural clarity, not measured performance.

Lite-Harness turns coding-agent use into an operations problem. It wraps Claude Code, Codex, OpenCode, and related tools behind one OpenCode-compatible API, then adds scheduled runs, secrets, persistent sessions, sandboxes, and approval routing.

The concrete example matters: an outreach agent runs every four hours on weekdays, stores vault keys, starts a test run, and asks for human approval before sending messages. The excerpt gives deployment mechanics and supported harnesses, but no reliability metrics or user study.

HermesBench and Dimensional Design focus on whether agent work can be checked in context. HermesBench scores a complete personal-agent configuration across 27 recipes, including tools, memory, safety, delegation, latency, and traces. Its public baseline is 78.2, with recipe definitions and redacted timelines available for inspection.

Dimensional Design adds a practical rule for AI-assisted work: put predictive steps behind deterministic pass-fail gates where exactness matters. Invoice totals, double-entry balances, plain-text collaboration, and small recorded human checks are the examples. This is guidance, not a benchmark, but it matches the period’s emphasis on visible limits.

Arch-Decision applies multi-agent coding work to a narrow team artifact: the Architecture Decision Record (ADR), a document that records why a software design choice was made. Its eight-phase process reads an issue, explores the codebase with three agents, proposes options, waits for approval, writes the ADR, and links it back.

OpenRouter supplies the market-side counterpart. Its gateway exposes more than 400 models and reports 8 million users plus about 100 trillion tokens per month. For agent builders, this points to model choice as a runtime dependency that can be swapped per task, while governance and review remain in the surrounding system.