Michael Rishi Forrester leads AI workforce transformation at Accenture LearnVantage. His teaching connects platform engineering with agent governance, emphasizing controls enforced by the surrounding infrastructure.[1]
Start with Build a Platform and Watch It Burn →
Links
Michael Rishi Forrester: speaking and biographyAI Workforce Transformation Lead, FDE and AI-Native Programs · Accenture LearnVantage
Career
Forrester’s professional biography describes nearly three decades across operations, cloud and platform engineering, including Red Hat, ThoughtWorks, AWS and KodeKloud. At Accenture LearnVantage, he works on forward-deployed engineer training and enterprise AI adoption. The current focus goes beyond teaching a tool: it includes the organizational and operating changes needed when agents begin carrying out engineering work. His public identity uses Rishi as his middle name.[1][2]
Contributions
In Build a Platform and Watch It Burn, Forrester and Whitney Lee let participants attack a deliberately permissive burrito-ordering agent connected to Kubernetes. The workshop contrasts prompt-based refusals with admission policy that can reject a deployment even when the model proposes it. It also examines input filtering, output filtering and traces of tool calls. The demonstrations expose different intervention points and their limits; a successful filter in one path does not prove that every route to the underlying action is blocked.[3]
Ideas
Agentic Covenants frames agent safety as a relationship between permission, responsibility and controls outside the model’s own instructions. The approach places architectural constraints around what an autonomous system can do. For a platform engineer, that means evaluating identity, network reach, deployment policy and evidence of actions alongside the quality of the generated answer.[4]
Tools, writing & talks
Writing, research & talks
Start here
Build a Platform and Watch It Burnai.engineerA live adversarial workshop comparing infrastructure policy, filtering and tool controls around an agent.
Watch the platform workshop first if you understand basic Kubernetes deployments. Follow one attempted action from the model’s proposal to the infrastructure decision; the traces help distinguish a refusal in chat from an operation that was actually prevented.
Co-presenter with Whitney Lee[3]A deliberately permissive burrito-ordering agent is attacked, then hardened with Kyverno policies and LLM Guard filters, while OpenTelemetry traces in Datadog expose prompts and tool calls.
Catalog workOriginal source- Agentic Covenantsmichaelrishiforrester.com
A governance framework for agent systems, centered on enforceable boundaries around autonomous actions.
Author[4]