The bill for skipping the architecture is coming due

This is a Decision Surfaces finding. When exception handling sits inside individual agents rather than at the orchestration layer, every edge case becomes an invisible cost center with no governance point. The study's three failure modes describe what happens when firms deploy agents faster than they redesign the decision architecture beneath them, a dynamic McKinsey's concurrent analysis confirms. Enterprise AI leads should audit their agent stack for ownership hierarchy before adding further platforms.

This signal maps directly to the Compiled Corporation lens. Automating decision execution without first mapping the decision architecture produces brittle systems. McKinsey's finding that process redesign lags agent deployment is the same structural gap the Fortune 500 cost study identifies from the bottom up. For Applied Identities clients, this is confirmation that Identity Architecture work precedes agent deployment, not follows it.

This is the Identity Control Surface signal of the week. Agent-to-agent communication is a new and largely ungoverned identity surface: when agents call other agents, the question of which identity is authorizing which action becomes both a security and a compliance problem. Zscaler's production deployments show that enterprises are beginning to treat this as infrastructure, not an afterthought. The semiconductor manufacturer case is especially instructive because it links a broad LLM rollout directly to agent communication governance.

Signed Agent Cards are the Identity Control Surface development that most enterprise AI programs have not yet acted on. Cryptographic identity at the protocol layer means that agent authentication is becoming a specification requirement, not a vendor option. Firms that have not mapped their agent identity architecture to the MCP 1.0 spec are accumulating governance debt. The A2A protocol's 150+ organizational support base signals that interoperability standards are consolidating faster than most enterprise security teams have planned for.

The Janus Brands tension is visible and widening. The same firms publicly advocating for slowdown controls are simultaneously scaling infrastructure, releasing frontier models, and expanding commercial deployments. For enterprise buyers, this creates a brand trust problem: safety messaging and product velocity are running in opposite directions. The DeepMind researcher's resignation adds internal credibility to the concern. Enterprises building governance programs around vendor safety commitments should weigh this signal carefully.

An agent that tests its own output is a Decision Surfaces milestone: the human review checkpoint moves further back in the loop, and the agent's self-evaluation becomes a governance surface in its own right. Enterprises evaluating AI-assisted software delivery need to define at what point in the test-deploy cycle a human decision is required, because Cognition's architecture is designed to operate without one. The cost reduction story from SWE-2 will accelerate adoption pressure on engineering organizations that have not yet answered that question.