Last issue closed on the forensic bill for an unverified edge. This weekend the industry converged on where the fix lives: not in the model, and not in the prompt, but in the explicit runtime wrapped around the model — the harness — and in the controls sitting on its edges. Stripe published the architecture of the agent 83% of its employees now use weekly. Microsoft moved its Agent Framework Harness to general availability. The EU switched on the law that governs every human-facing edge in an agent graph. And LangChain shipped both a benchmark for reviewer nodes and a gateway for spend and data controls. Five items, dates marked on each.
Stripe's Kai: 1,000 skills, one harness, 83% weekly adoption
Published today: LangChain's case study on Kai, Stripe's company-wide agent, built on the open-source Deep Agents framework. The numbers are striking — first version built in a week by one engineer, 296 to 5,000+ users in about four weeks, now 83% of Stripe using it weekly across 60,000+ sessions — but the architecture is the story. Four explicit layers (Deep Agents foundation, a Stripe-specific harness, a configuration layer, the UI); an S3-backed virtual filesystem so state survives across turns; a sandbox exposed as a tool rather than as the agent's runtime; and 1,000+ skills owned federally by 100+ teams, loaded through a two-pass gate because context degrades past ~150 skills. “The Deep Agents layer solves all the non-Stripey problems,” says Stripe's Sharadh Krishnamurthy. Note what the cap admits: context is a bounded resource, and skill selection is a conditional edge, not a convenience.
Microsoft's Agent Framework Harness reaches general availability
Reported this week by InfoQ: the Microsoft Agent Framework Harness and Foundry Hosted Agents are now GA, completing the framework's move from SDK to supported production runtime. The harness handles function invocation, multi-step execution, per-call history persistence, context compaction, tool-approval workflows, and built-in OpenTelemetry — plus connectors that let GitHub Copilot SDK and Claude Agent SDK agents compose inside one governed workflow. Principal engineer Wes Steyn's framing is the cleanest definition yet: a model alone only generates text; “you wrap it in a runtime; that runtime is the harness.” The same layer Stripe built bespoke, Microsoft now sells as infrastructure — the harness has become a product category.
The EU AI Act's transparency rules switched on August 2
As of Saturday, Article 50 of the EU AI Act is applicable law. Systems that interact with humans must disclose they are AI at the first interaction; providers generating synthetic text, images, audio, or video must mark outputs in machine-readable form; deepfakes and AI-written text published on matters of public interest must be disclosed; deployers of emotion-recognition systems must inform the people exposed to them. For agent builders the architectural translation is direct: every human-facing edge in your graph now carries a legally required disclosure property, and content-producing nodes need marking built into their output contract — not bolted on afterward. Human governance just moved from design principle to compliance requirement.
ReviewBench: code-review agents recover about 30% of what human reviewers caught
Published July 31: LangChain built a benchmark from real pull-request feedback in the LangSmith codebase — 59 tasks covering 64 issues that trusted human reviewers actually flagged — and scored agents on F1 across coverage and precision. Out of the box, current models recover roughly 30% of the baseline issues; structured review-strategy prompts improve scores materially. The uncomfortable implication for graph architects: if your verification gate is itself an LLM node, it has a measurable recall, and unmeasured critics are verification in name only. Evaluate the reviewer before you trust the gate.
LangSmith LLM Gateway puts spend caps and PII redaction at the model edge
Announced July 30: a governance layer that sits between agents and model providers, with time-bound spend limits at organization, workspace, API-key, and per-user level; model routing and failover across providers; rate limiting; PII redaction before data reaches a provider; and per-customer multi-tenant policies. This is Module 6's checklist — spend caps, permission boundaries, model routing — shipped as infrastructure rather than left as aspiration in a design doc. The pattern across this issue repeats: controls that used to live in prompts are migrating to explicit, inspectable layers on the graph's edges.