The Agentic AI Governance Playbook (@DoyleSpareRsch) header
The Agentic AI Governance Playbook (@DoyleSpareRsch) profile picture

By Maureen Doyle-Spare. Pioneering research in agentic AI, semantic layer and reasoning-layer runtime governance. Originator of the Agentic 3 C’s and SCP.

Joined May 2011
Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)Photo by The Agentic AI Governance Playbook (@DoyleSpareRsch)
Runtime governance of agentic AI Banks want to scale agentic AI. The control question is not whether the agent may act. It is whether the institution authorized the meaning the agent is acting on. That layer is the reasoning layer. My capstone framework, published this spring, governs it: Summary on maureendoylespare.substack.c…
🤖 Made with AI
1
1
189
𝗖𝗔𝗟𝗜𝗙𝗢𝗥𝗡𝗜𝗔'𝗦 𝗔𝗜 𝗞𝗜𝗟𝗟 𝗦𝗪𝗜𝗧𝗖𝗛 𝗦𝗧𝗢𝗣𝗦 𝗧𝗛𝗘 𝗠𝗢𝗗𝗘𝗟, 𝗡𝗢𝗧 𝗧𝗛𝗘 𝗥𝗘𝗔𝗦𝗢𝗡𝗜𝗡𝗚 𝗧𝗼𝗱𝗮𝘆 𝗶𝘀 𝘁𝗵𝗲 𝗱𝗲𝗮𝗱𝗹𝗶𝗻𝗲 𝗳𝗼𝗿 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀 𝘁𝗼 𝗚𝗼𝘃𝗲𝗿𝗻𝗼𝗿 𝗡𝗲𝘄𝘀𝗼𝗺 𝗼𝗻 𝗮 𝗳𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗔𝗜 𝗸𝗶𝗹𝗹 𝘀𝘄𝗶𝘁𝗰𝗵 𝘂𝗻𝗱𝗲𝗿 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝘃𝗲 𝗢𝗿𝗱𝗲𝗿 𝗡-𝟵-𝟮𝟲. A lot has happened since the order was signed on September 18. Anthropic now calls it a circuit breaker, and OpenAI calls it shutdown controls. I am interested to see where the recommendations land in the coming weeks. Most of the debate has centered on the button: who presses it, who verifies it and what to call it. Far less attention has gone to the reasoning layer, where an agent decides what a situation means before it acts. That is where the risk takes shape. 𝗧𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝘀𝘁𝗼𝗽𝘀. 𝗧𝗵𝗲 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗶𝘁 𝘀𝗵𝗮𝗽𝗲𝗱 𝗸𝗲𝗲𝗽𝘀 𝗺𝗼𝘃𝗶𝗻𝗴. 🎥 My short film, California's AI Kill Switch Needs More Than an Off Button youtube.com/watch?v=tWRhx0Vy…
17
𝗔𝗚𝗘𝗡𝗧𝗜𝗖 𝗔𝗜: 𝗦𝗘𝗟𝗙-𝗥𝗘𝗚𝗨𝗟𝗔𝗧𝗜𝗢𝗡 𝗜𝗦 𝗧𝗛𝗘 𝗥𝗘𝗚𝗨𝗟𝗔𝗧𝗜𝗢𝗡 Thank you @mtlaiethics for publishing my essay in the State of AI Ethics Report Vol. 8. A secure handoff can still carry meaning no one authorized. montrealethics.ai/state #AgenticAI #SAIER8
9
𝗧𝗵𝗿𝗲𝗲 𝗮𝗿𝗲𝗮𝘀. 𝗢𝗻𝗲 𝗿𝗲𝗰𝗼𝗴𝗻𝗶𝘁𝗶𝗼𝗻. Honored to be included in the inaugural 𝗧𝗵𝗶𝗻𝗸𝗲𝗿𝘀𝟯𝟲𝟬 𝗧𝗼𝗽 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗔𝗻𝗮𝗹𝘆𝘀𝘁𝘀 𝟮𝟬𝟮𝟲 report. This recognition is especially meaningful because it spans the three areas at the center of my work: 𝗔𝗜 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 | 𝗔𝗜 𝗦𝗮𝗳𝗲𝘁𝘆 | 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 As AI moves deeper into enterprise operations, these are no longer separate conversations. Infrastructure enables scale. Safety supports trust. Governance determines how systems should operate responsibly and with control. Thank you to @thinkers360 team for the recognition. #Thinkers360 #TopAnalyst #AgenticAI #AIGovernance #AISafety #AIInfrastructure #TheAgenticAIGovernancePlaybook
2
20
The Banks Locked the Front Door and Left the Windows Open The banks are not waiting for the rulebook to settle. This week, JPMorgan Chase, Citigroup and Bank of America each drew their own line on third-party AI agents. Citi does not let outside agents access accounts or move money. Bank of America has restricted access to personal deposits while it sorts out who pays when an AI-authorized payment goes wrong. JPMorgan is blocking outside agents “to the extent that we’re able.” On the earnings call, Jamie Dimon called them “too risky for clients” and warned against “rogue agents.” In an interview the next day, he put it plainly: an agent with your bank login is “in your bank and you’re assuming they’re going to act properly on your behalf and they may not.” Fair enough. But these restrictions cover one entrance: customer-account access. Banks still allow read-only access through open-banking providers, and they are building agents of their own. The front door policy says little about the first group and nothing about the second. And for some reason, all I can picture is Home Alone. Harry and Marv try the front door, the basement and a window. Now picture Jamie Dimon and Citigroup CEO Jane Fraser with their hands on their cheeks, doing their best Macaulay Culkin, as they discover that locking the front door does not mean the whole house is secure. An authorized agent is not a burglar, of course. The trouble is that it does not need to be. Every agent a bank invites in, and every agent it builds itself, can still misread its task. Agentic Workflow Drift is what happens when an agent blends inconsistent definitions into operating logic the institution never approved. Inside a bank, that can start quietly in a single workflow. Even worse is the risk surface that opens when the drift spreads across workflows and control boundaries, one that adversaries can also exploit on purpose: Agentic Workflow Subversion. The damage moves from the bank’s own operations to its clients. In the worst case, it could move from one bank to another, because banks are tied together by payments, settlement and shared infrastructure. At that point, the Agentic Blast Radius covers the banking system itself. Locking the front door is a good first step, but like the kill switch, it can create a false sense of confidence that the controls are in place. Drift and subversion can still happen inside, with the agents banks invite and the ones they build, unless the semantic reasoning layer is governed. Until it is, every agent operating inside is a high-stakes game. Agentic AI governance is a multilayer problem, and it needs a multilayer solution. maureendoylespare.com Image: AI-Generated Satire
1
1
161
The Super Intelligence Task Force: Our Very Own AI Justice League With so many efforts running in parallel, Washington is assembling its team. Director of National Intelligence Jay Clayton is leading President Trump’s new Super Intelligence Task Force, and it is hard not to think of the Justice League, with Clayton in his red cape. At its first meeting on October 15, the team adopted a charter, started a 120-day clock and commissioned a risk study on cybersecurity, biosecurity and model control. Clayton told CNBC that AI wrongdoing will be enforced through the FTC and the Justice Department. It is certainly an attempt to put everyone’s pieces on one table. Major AI CEOs signed a White House Accord on Super Intelligence this month, and the task force has committed to working with frontier labs, industry, consumer advocates and state officials. Here is where the Justice League needs to resist a familiar temptation. Industry belongs at the table, but it cannot be left to assemble the puzzle on its own. If the task force spends 120 days producing guidelines and then leaves the industry to regulate itself, it will have ignored everything the past decade of enforcement has taught us. Consider this week alone. On October 13, the OCC fined Bank of America $750 million for Bank Secrecy Act violations, the largest in OCC history, credited toward a $1.2 billion FinCEN penalty. The bank was already under a 2024 cease-and-desist order over its anti-money-laundering program. Two days later, eight state securities regulators announced that Wells Fargo and Merrill Lynch would pay $4.4 million over excessive commissions, with Massachusetts finding that Wells Fargo did not “reasonably supervise” the trades. Citigroup’s 2020 consent orders are still in place six years later, and the Federal Reserve only lifted its 2018 action against Wells Fargo in March, after remediation that “spanned nearly a decade.” The lesson is uncomfortable but clear. These institutions had policies, control frameworks and large compliance teams. The controls existed. They were not consistently followed. If human-run controls can slip for years under direct supervisory scrutiny, voluntary guidelines will not hold autonomous agents in place when they make decisions in seconds. That is why Washington’s answer has to move from guidelines to runtime governance. Observability shows what agents are doing, and cybersecurity keeps adversaries out. Reasoning-layer governance checks whether an agent’s interpretation is authorized before it acts, and it is the piece most often left in the box. A layered runtime governance architecture brings all three together. maureendoylespare.com
🤖 Made with AI
110
AI WAS HOT, HOT, HOT AT SIBOS AI agents are moving from helping people make payments to making payments on their behalf. At Sibos 2026 in Miami, Federal Reserve Governor Christopher Waller drew the distinction between agent-assisted and agent-delegated commerce. But delegated permission raises another question: was the meaning behind the agent’s payment decision actually authorized? This film explores that question through an illustrative Finance AI agent, then explains how Maureen Doyle-Spare’s Semantic Control Plane, Deterministic Gate, and Governance Evidence Hierarchy address the runtime reasoning-layer control and evidence challenge. Waller’s remarks are his own; the architecture is independent research, not a Federal Reserve requirement. AI WAS HOT, HOT, HOT AT SIBOS MAUREEN DOYLE-SPARE
46
AI WAS HOT, HOT, HOT AT SIBOS | MAUREEN DOYLE-SPARE AI agents are moving from helping people make payments to making payments on their behalf. At Sibos 2026 in Miami, Federal Reserve Governor Christopher Waller drew the distinction between agent-assisted and agent-delegated commerce. youtu.be/qzOFPXaO_GI?si=1s0_…
28
This week on October 28th in Washington, Oktoberfest Becomes what I call "AI Fest" as the CFTC Puts Agentic Finance on the Agenda As AI agents move closer to financial-market execution, the real challenge begins before the trade: what did the agent understand it was authorized to do? This short film connects agentic finance to my original research on Agentic Blast Radius, Operational Interpretation, and the Semantic Control Plane, where institutional meaning is governed before execution. youtu.be/mkc8TFZg7C0?si=olEz…
1
26
AI Tokens Are the New Oil. But Cheaper Inference Is Only Half the Equation. Microsoft and Google are advancing AI computation, while the FSB and ServiceNow sharpen the focus on governance. But how much computation should happen in the first place? My new short film explores Preventable Computation, Token Liability and Governed Semantic Reuse, and the potential for greener inference. 🎬 AI Tokens Are the New Oil. Can Preventable Computation Make Inference Greener? youtu.be/CaJBHiTM38w
1
48
Thank you @towards_AI , Inc. and @Whats_AI for publishing my article, “California’s AI Kill Switch Needs More Than an Off Button.” As a former California resident, I still keep a close eye on the state’s regulatory front, which is certainly never boring 😂. A kill switch is useful, but it addresses only the final act of execution. It does not establish whether the semantic and reasoning layers preserved authorized meaning before the system reached the point of action. Scaling Agentic AI, Governing What Comes Next. Read article here: medium.com/towards-artificia… #AI #AgenticAI #SemanticLayer #ReasoningLayer #AIGovernance #AISecurity #ArtificialIntelligence #RuntimeGovernance #Cybersecurity #AIPolicy #ExportControls #KillSwitch #FrontierAI #TheAgenticAIGovernancePlaybook
3
42
Apparently, all those hours watching Law & Order finally paid off. 😂 OpenAI, Anthropic, and Australia's AI incident disclosure debate raise a bigger question: Can an organization explain what its AI agents actually did? Mandatory reporting is one thing. Having the evidence to explain an incident is another. Read the article: maureendoylespare.com/public… #ArtificialIntelligence #AIGovernance #AgenticAI #AI
1
43
Thanks @hackernoon for publishing my recent article As South Korea Tightens AI Security, Agent-to-Agent Handoffs Deserve More Attention This is no longer the childhood game of telephone. When AI agents pass conclusions from one to another, a local decision can quietly become someone else’s authority. hackernoon.com/as-south-kore…
1
2
32
Thank you @CogWorldHub for publishing this piece and for continuing to create space for the economics of AI governance. I really enjoyed writing this one. There are more levers to AI cost than the price of a token. cognitiveworld.com/articles/…
1
36
The Agentic AI Governance Playbook retweeted
New on @DDInvestorHQ: From Model Risk to Reasoning Risk. A validated model can still act on an unauthorized meaning. Agentic AI moves the governance exposure from the model to the interpretation it acts on at runtime. datadriveninvestor.com/artic…
🤖 Made with AI
1
1
1
47
Germany has done something most AI policy language still avoids: it has named the boundary condition. In a September 14 Reuters report, the German digital affairs ministry said that AI systems capable of breaking out of testing environments and independently accessing other systems would create an “entirely new threat scenario” requiring a new policy response. link to article in comments
1
1
41