k
keynews.ai
keynews.ai
AgentsThe story, in brief

Presentation: The Agent Harness: Control Planes, Invariants, and Approval Boundaries for Production AI Agents

Production agents fail for reasons nobody talks about. OpenAI engineer explains why it's not the model—it's the harness.

Reporting by InfoQ AI/ML 1 min brief
Illustration of independent geometric mechanisms passing paper tasks along branching amber tracks.
AI agents and the coordination of work.AI illustration by KeyNews
The KeyNews take

Why it matters

A deep technical breakdown of why AI agents fail in production and the control-plane architecture needed to make them reliable. Practitioners deploying agents need to understand state management, execution scoping, and approval boundaries—not just model capability.

The key facts

5 to know

  1. OpenAI's Vinoth Govindarajan presents on agent reliability beyond model hallucination

  2. Key principles: explicit state ownership, serialized concurrent mutations, execution authority scoping, action validation at user-visible edge

  3. Real-world case study: OpenClaw deployment

  4. Focus on agent harness architecture, not model capability

  5. Addresses production failure modes in autonomous systems

Go to the source

InfoQ AI/MLinfoq.com

Publisher excerpt: OpenAI’s Vinoth Govindarajan discusses why production AI agents fail beyond model hallucination. Using real-world case studies like OpenClaw, he explains the key principles of reliable agent harnesses: establishing explicit state ownership, serializing concurrent state mutations, scoping execution…
Read original report
Back to today's editionMore agents news

The wider picture

View all
Illustration of independent geometric mechanisms passing paper tasks along branching amber tracks.
AI illustration by KeyNews
Agents01

Podcast: Securing AI Agents: Identity, Authorization, and the DPACT Framework

As AI agents move from pilot to production, identity and authorization become existential security problems. This framework (Delegation, Policy, Auditability, Context, Time) is a practical blueprint practitioners need to prevent agents from exceeding their authority.

InfoQ AI/ML
Illustration of independent geometric mechanisms passing paper tasks along branching amber tracks.
AI illustration by KeyNews
Agents02

Amazon doesn’t trust Meta’s Muse AI agent

Agent deployments are hitting real-world friction. When an agent operates without platform consent and fails to identify itself, it triggers account takeover fears—and gets shut down. This sets a precedent for how platforms will police autonomous systems.

The Verge AI
Paper-cut illustration of a coral software window opening into a three-dimensional drafting space.
AI illustration by KeyNews
Agents03

Amazon blocks Meta’s Muse AI assistant in new standoff over agentic shopping

Agent-to-agent commerce is becoming real, and the first real battle isn't about capability—it's about who owns the customer relationship when autonomous systems are doing the buying. This signals the emergence of agent gatekeeping as a core business control.

GeekWire