Nate Jones told a story last week that explains the whole quarter.
He runs six calendars. An important milestone was the day an agent could read all six and identify a free moment. An answer built from two calendars would sound perfectly reasonable, and it would be completely wrong.
That is the actual risk today. It's not the rogue agent; it's the agent giving a reasonable answer from half the context, while nothing in the stack can tell you which half.
Consider that alongside Nate's arithmetic:
Individual implementation speed: 10x
Realized business speed: ~1.8x
The rest evaporates between the fast person and the rest of the organization: code arriving faster than review, prototypes landing ahead of decisions, and context that never leaves an isolated session.
Over the last 90 days, the AI frontier moved aggressively toward the request path. Almost none of it moved toward the record.
State of the Frontier: Q3 2026
1. Regulators paused the timeline, but sharpened the record
EU AI Act: Regulation (EU) 2026/1744 pushed high-risk deadlines to December 2027, but left Article 12 untouched: systems must technically allow automatic lifecycle recording, with a six-month retention floor (Articles 19 & 26).
US Banking Regulators: The OCC, Fed, and FDIC rescinded SR 11-7, excluding generative/agentic models from traditional model risk guidance. Meanwhile, FINRA urged firms to retain prompt/output chains for multi-step reasoning auditability.
States & Standards: Colorado's SB 26-189 locked in a three-year retention mandate. The federal Stop Rogue AI Act directed NIST toward agent identity and kill-switch standards. Five Eyes guidance demanded per-agent cryptographic identity and complete tool-call logging.
The convergence: Nobody cares which model you choose. Everyone is asking what the agent knew, what it executed, and whether you can prove it in court.
2. Identity and the request path consolidated
Identity as table stakes: Microsoft Entra Agent ID went GA; Okta announced Agent SSO, agent-to-agent peering, and an upcoming runtime kill switch. In parallel, platform M&A wiped out standalone non-human identity (CrowdStrike/SGNL, Cisco/Astrix, SailPoint/Entro, Cyera/Oasis, Okta/Permiso).
The battle for the transit line: Palo Alto Networks acquired Portkey to own the AI gateway. Stripe acquired OpenRouter to capture token settlement. Kong shipped AI Gateway 2.0 with MCP support.
The blind spot: Microsoft's MCP firewall covers remote HTTP servers, not the thousands of unmanaged stdio servers running locally on developer machines.
3. Analysts mapped the stack, and left the core empty
Gartner's new AI Governance MQ requires runtime enforcement, noting 80% of unauthorized agent transactions through 2028 will stem from internal policy violations, not external attacks.
Forrester notes that logging autonomous actions for auditors carries an unsustainable post-hoc cost.
Context engineering peaked on the Hype Cycle; context graphs sit at under 1% penetration. As Atlan framed it: Context is produced by every function and owned by none.
Three Directions Worth Betting On (And Their Hidden Tradeoffs)
1. Identity on Every Surface
The bet: Issue, scope, and revoke credentials for every copilot, tool, and MCP server.
The vendor tradeoff: Identity stops at the door. It tells you who acted, not what they knew or which calendars they checked.
2. Inline Runtime Control
The bet: Intercept and enforce policy in the flow rather than observing post-hoc dashboards.
The vendor tradeoff: The transit tax. Whoever hosts the gateway cloud accumulates your operational traffic and IP.
3. The Record as Evidence
The bet: Retain verifiable, auditable chains matching human principal, surface, context, and tool calls.
The vendor tradeoff: High cost if bolted on. Post-hoc reconstruction is expensive; it only works if captured directly in the flow.
The Tradeoffs Nobody Puts on the Slide
Speed vs. The Record: Inline inspection adds single-digit millisecond latency. Skipping it accepts total liability when an un-remediable action executes.
Sovereignty vs. Convenience: Sovereign AI is simple: the record of what your surfaces did lives on infrastructure you control, under keys you hold, and leaves only when you export it.
Watching vs. Enforcing: Post-hoc observability lets un-enforced policy violations run silently for quarters.
Cost of the Record vs. Cost of Absence: Nate's 10x-to-1.8x gap is the cost of absence: the review tax, stale context rework, and re-bought token compute.
The missing primitive is memory. A reasonable answer derived from half the context is neither a traditional security breach nor a model failure. It is a memory failure. Models hold static weights; compounding business learning lives entirely in runtime context. Whoever controls context controls compounding.
Where a Memory Layer Sits
Excipio deploys inside the enterprise perimeter, directly in the request path on your infrastructure:
Consumes Identity: It doesn't replace Entra or Okta. It records what the credential did, who it talked to, and what was in context when it fired.
Keeps the Path Sovereign: Transit and accumulation stay on your cluster. Within 30 days, the telemetry highlights overlapping context, stale reads, and redundant token spend.
Builds Graphs from Telemetry: The knowledge graph mirrors actual runtime traffic, ensuring the memory regulators ask for is the exact memory the enterprise used.
The industry spent the quarter buying the front door and the highway. Regulators are asking for the record of the trip.
If that record lives in a vendor's dashboard, it is marketing. If it lives inside your perimeter, it is evidence.
Primary Sources
Regulation (EU) 2026/1744 (OJ, July 24, 2026); EU AI Act Arts. 12, 19, 26
US Interagency Guidance: OCC Release 2026-29; FINRA Oversight Report (2026); FSB Sound Practices Consultation
Statutes & Bills: Colorado SB 26-189; Stop Rogue AI Act (H.R. Sept 2026); CISA/Five Eyes Agentic AI Guidance
Industry Telemetry: OWASP LLM Top 10 (2026); Gartner AI Governance MQ & Guardian Agent Guide; Forrester Agentic AI Study; Cohere Sovereign AI Report; Nate Jones (Nate's Substack, Sept 2026)