Episode Details
Back to Episodes
AI Agents Architecture: The Secret Architecture That Makes AI Agents Actually Work
Season 1
Published 4 months, 3 weeks ago
Description
(00:00:00) The Validator's Triple Check
(00:00:07) Capability, Policy, and Feasibility: The Validator's Three Pillars
(00:01:47) The Triogate: Ensuring Safe Execution
(00:02:59) Implementation and Architecture
(00:04:19) Subscribe and Watch Next Episode
(00:04:36) The Executor's Role: Operations and Guarantees
(00:08:41) Workflows as Graphs: Structuring Reliability
(00:12:16) Observability and Security in Graph Validation
(00:12:53) Microsoft 365 Integration: A Secure Architecture
(00:22:31) Measuring Success: Metrics and Benefits
In this episode of M365.fm, Mirko Peters explains why most AI agents don’t fail because the prompt is bad — they fail because there is no real architecture behind them. You’ll see how separating cognition (LLMs) from operations (executors), plus adding validation and explicit workflows, turns “smart but flaky” agents into stable, predictable systems that enterprises can actually trust.
WHAT YOU WILL LEARN
Prompts are thoughts. Executors are actions. Validation is safety. When you rely only on prompts, the model hallucinates tools, ignores preconditions, and happily produces “partial success” that breaks downstream systems without throwing an error. The fix is a contract‑first design: each node in a workflow has explicit inputs, outputs, and postconditions, and every tool call is checked against a policy and schema before it runs.Mirko shows how this looks in practice: DAG‑shaped workflows with clear state boundaries, compensation logic for side effects, and node‑level tracing so you can replay exactly what happened. Static validation catches cycles, unreachable nodes, and broken contracts before deployment; runtime guards enforce RBAC, ABAC, scopes, and safe egress. With Microsoft Graph as the grounded data layer and Azure OpenAI as the reasoning engine, th
(00:00:07) Capability, Policy, and Feasibility: The Validator's Three Pillars
(00:01:47) The Triogate: Ensuring Safe Execution
(00:02:59) Implementation and Architecture
(00:04:19) Subscribe and Watch Next Episode
(00:04:36) The Executor's Role: Operations and Guarantees
(00:08:41) Workflows as Graphs: Structuring Reliability
(00:12:16) Observability and Security in Graph Validation
(00:12:53) Microsoft 365 Integration: A Secure Architecture
(00:22:31) Measuring Success: Metrics and Benefits
In this episode of M365.fm, Mirko Peters explains why most AI agents don’t fail because the prompt is bad — they fail because there is no real architecture behind them. You’ll see how separating cognition (LLMs) from operations (executors), plus adding validation and explicit workflows, turns “smart but flaky” agents into stable, predictable systems that enterprises can actually trust.
WHAT YOU WILL LEARN
- Why prompts alone can’t guarantee correct, repeatable behavior in real workflows
- The difference between thinking (LLM) and doing (executors with contracts, retries, and postconditions)
- How workflow graphs (nodes, edges, state, compensations) give agents a real map instead of improvisation
- How static graph validation and runtime policy checks catch bad plans before they hit production systems
- How to use Microsoft 365 Graph as a grounded data layer with least‑privilege access and citations
- How Azure OpenAI, schema‑bound outputs, and Copilot Studio orchestration fit together in one stack
- Which metrics actually prove that your agent is reliable: accuracy, p95 latency, cost, and first‑pass completion
Prompts are thoughts. Executors are actions. Validation is safety. When you rely only on prompts, the model hallucinates tools, ignores preconditions, and happily produces “partial success” that breaks downstream systems without throwing an error. The fix is a contract‑first design: each node in a workflow has explicit inputs, outputs, and postconditions, and every tool call is checked against a policy and schema before it runs.Mirko shows how this looks in practice: DAG‑shaped workflows with clear state boundaries, compensation logic for side effects, and node‑level tracing so you can replay exactly what happened. Static validation catches cycles, unreachable nodes, and broken contracts before deployment; runtime guards enforce RBAC, ABAC, scopes, and safe egress. With Microsoft Graph as the grounded data layer and Azure OpenAI as the reasoning engine, th