Reading on governing AI agents, and where each piece meets the map.

Identifiers, real-time monitoring, and activity logs as the three levers of visibility; the set’s e1, p5, and w5, with the mandate added as the fourth.

The systems outside the agent that attribute, shape, and remedy its actions; the control plane is one such infrastructure, inside one firm.

The parties of the agent lifecycle and baseline practices for each; the set assigns those practices to boxes with owners.

Agents read through agency law and the principal-agent problem; the mandate on the map is a principal’s instruction made machine-readable.

Agent benchmarks ignore cost and reproducibility; the set’s evaluations at p6 are run by a team that did not build the agent, with cost beside accuracy.

On a SWE-bench Verified subset the harness explained 7.8 times more variance than the model; so the firm approves a release, never a model (part three).

Agency as a matter of degree, with harms that grow as goals get less specified and impact more direct; why autonomy is granted level by level.

Autonomy, efficacy, goal complexity, and generality as the dimensions of an agent; the labels of the ladder A.0 to A.5.

The risks to people rise with each level of autonomy; A.5 is the level no evidence condition allows.

Why an agent can be right at every step and wrong as a whole: the failure the loop, the current-state input of the check, and situation card 5 are built against.

Knowing, saying, and doing as layers of governance, and a circuit breaker that stops an act on a basis below standing: the check at w4.

How human-agent loops build software, with verified outcomes, auditable evidence, and verification gates: the release at w1, the harness at p1, evaluations at p6.

Scoped, traceable authority granted by a person to an agent, on OAuth and OpenID; the set’s identity at p1 and entitlements at p3, under the mandate at w3.

Existing identity and authorization standards applied to agents; the platform row of the map in a standards body’s words.

Five overlays on SP 800-53, two of them for single-agent and multi-agent systems; the platform controls the set leaves to the CISO.

Instructions planted in what an agent retrieves take over what it does; why the check reads the act and its basis itself, whatever the agent declares.

Prompt injection and excessive agency at the head of the practitioners’ list (LLM01 and LLM06); the guardrails at p4 and the entitlements at p3.

Third-party dependence, concentration, and cyber risk as the system-level questions; the set stays inside one firm and names those as the layer above it.

Governance, model risk, data, and third parties as the gaps supervisors see; four of the set’s essentials, with the mandate and the loop missing from the list.

How existing insurance rules apply to AI, proportionately; the row of part four’s law slide that the motor case reads.

The AI Act’s high-risk rules set against banking law, with no significant contradiction found; the ground for the credit cases’ law rows.

Securities supervisors on AI in trading, advice, and operations, and the risks they see; the ground for the capital-markets case, beside the EBA’s view for banks.

Firm-level benefits and risks and how they become systemic; the firm-level half is what the map governs.

Model identification, governance, development, validation, and risk mitigants; model risk validates the release at w1 and the evaluations at p6.

A risk-based approach to model risk that leaves generative and agentic AI to the bank’s own risk management; the register quotes the footnote (row 10).

Govern, map, measure, and manage as the functions of AI risk management; the enterprise row of the map in a framework’s words.

A management system for AI that can be certified; the owners, the register, and the reviews on the map, as an auditor of the standard would sample them.

An end-to-end framework for internal algorithmic audit; the evidence base at e5 and the record at w5 are what such an audit samples.

Governance, model, and application audits as three layers; the map’s three rows, audited in turn.

Entities, activities, and agents as the grammar of provenance; the first thing a claim carries, in a standard’s terms.

Retrieval paired with generation; what p2 does, and what the set says it cannot do alone: find the version that applies.

Models read the start and the end of a long input and lose the middle; why the agent receives a bounded package of claims and never the library.

A claim checked against evidence that supports or refutes it; the claim grain, in the literature where it was first worked at scale.

Whether a generated statement is attributable to an identified source; the test the record at w5 answers for every premise of an act.

Why fluent output is no evidence; the record keeps the claims an act relied on, and never the model’s prose as a source.

Language models and knowledge graphs combined; the knowledge estate at e2 as structured, versioned claims beside the model.

The lifecycle that keeps knowledge fit to act on, the decay classes, and why a model’s own output never re-enters as an independent source: chapter 05, how knowledge is prepared.

The claim and its standing on two axes, admissibility computed per act against the consequence tier’s bar, and the signed manifest: the set’s second rule.

The type system beneath the claims, and how it stays versionable and transferable between domains.

Governed claims over stored documents, traceable provenance over trusted sources, and preserved contradictions over forced consensus: the values behind chapter 05.

Coordination as the management of dependencies between activities; what a cohort of agents needs once no single mandate holds the limit.

Insects coordinating through the traces their work leaves behind; the record at w5 and the knowledge at w2 as the medium agents read.

Miscoordination, conflict, and collusion among agents as distinct risks; the cohorts and cascades the map does not draw yet.

Navigation on a ship as cognition spread across people, instruments, and records; the record and the register as part of how a firm thinks.

Organisational intelligence as map, physics, and player: the frame around the knowledge plane.

Intelligence as a property of the medium the units coordinate through: what the loop at w6 feeds, and what lies above one act.

Accountability scoped by consequence class, and containment as multi-agent: e1, e4, p4, and w6, and what the map does not draw yet.

Agents that plan by predicting the consequences of their acts in a learned world model; the fifth card of what the map does not draw yet.

I-JEPA, the first of the line; a predicted consequence enters the check as one more claim with a standing.

V-JEPA, the same idea on video; the map does not move when the architecture does.

Four stages and ten levels of automation; the ancestor of every autonomy ladder, the set’s A.0 to A.5 included (A.5 as the end no evidence condition allows).

The more a system is automated, the more skilled and less practised its supervisor; the people slide of part four, forty years early.

Reliance calibrated to what automation can actually do; the standing an act needs is set by its consequence, never by the agent’s confidence.

Control that tracks human reasons and traces to a responsible person; the mandate with a named owner, and the record that names who decided.

Forty-one oversight policies that ask a person to catch what a person cannot; why the set makes human in the loop two jobs, deciding referrals and reviewing samples.