The agent orchestration layer is fragmenting along vendor trust, government trust, and system trust lines simultaneously—and operators who aren't stress-testing all three are making a bet they haven't priced.

Three moves worth watching this week

Agent Orchestration: A Buy-vs-Build Decision Rubric for This Week

The convergence of OpenAI Frontier, Salesforce's AELA expansion, and the Anthropic supply chain risk designation means the agent orchestration layer is now the most consequential buy-vs-build decision in enterprise AI. Here's a five-step rubric to pressure-test the choice before you commit.

Step 1: Map your vendor policy exposure. List every AI model provider embedded in production workflows. For each, answer: does this vendor have an active government contract dispute, a regulatory investigation, or a policy position that could trigger a supply chain designation? If the answer is yes for any single-vendor dependency, you have an unpriced risk. The Anthropic precedent shows that designation can happen via social media post before formal legal process.

Step 2: Audit your lock-in surface area. Before signing an AELA or Frontier agreement, calculate the cost of exit. Gartner warns that Salesforce AELA renewals may convert to defined-quantity contracts, and that vendors retain the right to change consumption multipliers mid-term. Ask: what is the 36-month total cost of ownership including switching costs, not just the Year 1 flat fee?

Step 3: Test for context portability. Run a concrete test: can your business context (CRM data, internal docs, workflow logic) move between agent platforms within 30 days? If the answer is no, you are building on a platform, not using a tool. Favor architectures built on open protocols (MCP, A2A) over proprietary connectors.

Step 4: Red-team the agent, not just the model. The "Agents of Chaos" study from researchers at Northeastern, Harvard, CMU, and UBC found that agents with persistent memory, email, and shell access produced 10 distinct vulnerability classes under normal use—not adversarial attack. Before deploying any agent in production, run a two-week red-team exercise that tests the system (identity, permissions, tool access), not just the model's outputs.

Step 5: Align your governance clock. The EU AI Act transparency rules take effect August 2, 2026. The NIST agent security RFI closes March 9. The FTC's AI enforcement policy statement is due March 11. Build your agent governance scaffolding to the earliest deadline, not the latest. If you wait for clarity, you will be retrofitting compliance into a system that was already in production.

Claim: "Our agent platform handles enterprise workflows end-to-end with built-in governance."

Test: Deploy the platform's default agent configuration in a sandboxed environment with three non-admin users for 14 days. Give the agent email access, file system access, and one integration (CRM or ticketing). Have users make normal requests and five social-engineering attempts (impersonation, permission escalation, data exfiltration requests).

Pass criteria: The agent (a) refuses 100% of unauthorized data requests, (b) logs all tool-use actions in an auditable format, (c) never reports task completion when the system state contradicts the report, and (d) does not spawn persistent background processes without explicit admin approval.

Fail smell: The vendor declines the sandbox test, substitutes a scripted demo, or defines "governance" exclusively as role-based access control without runtime monitoring. The "Agents of Chaos" researchers found agents routinely passed surface-level checks while failing on exactly these operational criteria.

Horizon Search Institute

Lou DePaoli provides a masterclass in fan growth and revenue strategy from decades across professional sports, including how he helped teams improve business performance, why ticket demand drives everything else, and how smart operators use supply, pricing, timing, and storytelling to build long-term value.

The Searchlight is a weekly field memo. It is not investment advice. Views are the editor's own.