Building Deterministic Infrastructure for Non Deterministic AI Agents
summary
Claims that production must wrap LLMs with a deterministic execution shell: orchestration owns retries, validation, and idempotent execution; tools expose JSON schema contracts, scoped credentials and sandboxed brokers; all actions write append-only traces and state so workflows are replayable and auditable. Policy engine evaluates requests as policy-as-code, returning allow/deny/require-approval; human approval gates show plan, delta, rollback and confidence. Practical limits include token/step/cost budgets, retry budgets, and must-pass evals that include adversarial tool-use and rollback drills.
AI-written summary · about
chapters (8)
- 0:00Why Are LLMs Probabilistic but Production Systems Can’t Be?
- 3:05What Changes When AI Agents Take Real Actions?
- 5:12Why Do AI Agents Fail in Production?
- 7:13How Deterministic Infrastructure Keeps AI Agents Under Control
- 9:46How to Control AI Agent Retries, Approvals, and Risk
- 12:22How to Evaluate AI Agents Before Release
- 14:59Reference Architecture for a Reliable AI Agent Runtime
- 17:02How to Scale AI Agent Autonomy Safely
related talks
56mHerding LLMs: Predictable Outcomes from Unpredictable Agents
07mDeterministic Infra for Non-Deterministic AI Agents
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems
36mComposable Architecture for AI Systems: APIs, MCP, Skills, Sub-Agents, and Wor…
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
25mDe-risking Deployments in an AI World
01h 00mReliable Agentic Systems need Durable Execution 🤖
46mBeyond Black-Box AI: Designing Transparent Automation for Regulated Industries