Agentic AI Governance

8/5/2026

This post extends the agentic AI governance thread by focusing on the practical application of agentic AI by leveraging existing predictive AI foundations. It introduces the concept of 'agent workforces' as the most impactful category of agentic AI, emphasizing the orchestration of data, models, and applications around defined business outcomes. The post provides strategic guidance on transitioning to agentic AI, including considerations for model strategy, cost optimization (value vs. tokens), and outcome-driven development. It also includes real-world examples from Chevron and a technology company to illustrate the application of agentic AI in operational scenarios.

7/29/2026

This post defines a 30-day practical checklist for implementing agentic AI governance. It breaks down the process into five key phases: identifying agents and their authority, enforcing runtime permissions, governing high-impact actions outside the model, enabling reconstruction of consequential actions, and ensuring safe failure modes. It emphasizes the need for deterministic controls, least-privilege access, comprehensive logging, and defined incident response procedures, establishing a foundational framework for managing the risks associated with autonomous AI agents.

7/24/2026

Introduces the concept of agent identity as a lifecycle, moving beyond static settings to dynamic management. Proposes just-in-time (JIT) credential provisioning and programmatic revocation as key mechanisms to eliminate standing privilege. Highlights the operational feasibility of JIT access for agents compared to humans and references OpenID Provider Commands for identity lifecycle operations (activate, suspend, revoke, delete). Emphasizes continuous runtime authorization evaluation tied to the agent's task.

7/22/2026

This post introduces the concept of federating trust outward from a native governance plane to existing identity providers and workload identity systems, addressing the challenge of organizational heterogeneity. It discusses the use of standards like SPIFFE, WIMSE, and OIDC for composing solutions across trust domains. The post also emphasizes the need to scale controls based on an agent's blast radius and identifies cross-trust domain interaction as an open problem.

7/20/2026

Introduces the concept of credential isolation as a security measure for agentic AI systems. It proposes a 'broker' pattern where a mediating service holds sensitive credentials and enforces access scope for agent requests, preventing raw credentials from entering the model process. This mitigates risks associated with prompt injection leading to credential exfiltration. The post also emphasizes the importance of controlling egress paths and failing closed to ensure the broker pattern is consistently applied.

7/17/2026

Introduces DataRobot OpenCode, a coding agent that integrates with the DataRobot LLM Gateway. Details installation via DataRobot CLI and selection of models from the gateway. Highlights pre-installed DataRobot agent skills like Agent Assist.

7/15/2026

This post extends the 'Agentic AI Governance' feature thread by detailing the infrastructure required to scale AI agent governance from a few agents to hundreds. It elaborates on the challenges of manual governance breaking down at scale, the necessity of centralized controls, and the specific components of governance infrastructure: agent registry, identity, policy propagation, permission scope, tool access, component lineage, runtime enforcement, monitoring, audit trails, and review triggers. It specifically unpacks the importance of centralized agent identity, policy propagation, and cross-environment enforcement as critical differentiators for large-scale governance.

7/14/2026

This post defines the four essential components of a first-class agent identity: a distinct principal, scoped permissions, a clear owner, and an independent kill switch. It contrasts this with common alternatives like shared service accounts, per-user impersonation, and static secrets, highlighting their shortcomings. The post further elaborates on the conditions under which agent identity collapses into workload identity (one-to-one mapping, registry as source of truth, identity continuity) and when it requires a separate layer, citing examples from Microsoft and AWS. It sets the stage for future discussions on identity propagation across hops and trust domains.

7/10/2026

This post identifies and elaborates on the critical security and operational risks associated with AI agents inheriting human credentials, leading to privilege escalation, lack of scoping, poor attribution, and difficult revocation. It argues for a shift from a 'passport' model of identity to an 'authority' model where permissions are determined at runtime based on the agent's specific task and context, addressing the non-deterministic nature of agents. It also critiques the structural assumptions of traditional IAM stacks that do not adequately categorize or govern agents.

7/8/2026

This post introduces the concept of agents as a third type of actor in identity and access management (IAM) systems, distinct from human users and service accounts. It highlights the non-deterministic nature of agents and the resulting security gap where agents borrow human credentials, leading to privilege escalation and auditability issues. The post outlines immediate checks for borrowed human API keys, audit log separation, and agent revocation without collateral damage. It also sets the stage for a series that will detail building a robust agent identity and governance system, including first-class agent identity, delegation chains, credential management, and lifecycle governance.

7/1/2026

Introduces syftr, an open-source tool for optimizing agent workflow design using multi-objective Bayesian optimization to find Pareto-optimal configurations, significantly reducing search compute and cost. Also introduces Token Pool, a Kubernetes-native API gateway layer that manages multi-tenant inference capacity using inference-native units (token throughput, KV cache, concurrency) and a debt-based fairness mechanism to prevent service degradation under load. The post also discusses the historical evolution of DataRobot's open-source contributions, from predictive AI to agent infrastructure, and outlines the next steps in connecting production telemetry to the optimization layer.