Building the future of agentic infrastructure

The conversation explores the rapid evolution of agentic infrastructure over the past six months, moving beyond basic API access to inference toward rich platform features that reduce infrastructure and harness engineering burdens. The speakers emphasize that modern models' improved capabilities have reduced the need for excessive scaffolding around agent behavior, allowing agents to make intelligent decisions within reasonable guardrails.

A central theme is agent identity and governance. Rather than treating agents as black boxes, the vision is for agents to have distinct identities that request specific permissions for tasks. When an agent needs access to resources A, B, C, and D to accomplish an outcome, users can selectively grant permissions and audit agent actions. This creates accountability while enabling autonomous work.

The discussion covers how agents can interoperate through APIs, with examples of developers building thin MTP servers to expose agent capabilities to other agents. The speakers note that harness layers (the control structures around agents) are becoming thinner as models improve, though new meta-strategies are emerging—including competitive multi-agent approaches, adversarial pairs, and advisor strategies where agents consult with peers.

Real-world applications demonstrate significant complexity being solved. One hackathon winner created a system to replace retiring factory experts by combining monitoring signals, standard operating procedures, and agent decision-making. Engineering teams are building end-to-end development systems using agents for code writing, testing, documentation, and QA—going beyond simple code generation to full project orchestration (exemplified by Shopify's River).

Barriers to adoption include security and compliance frameworks designed for pre-agent assumptions, and the challenge of evaluating agent performance. The speakers advocate for ROI calculations starting with individual productivity gains, scaling to team-level metrics, and only then addressing company-wide process transformation. They warn against over-optimism creating false independence across teams without coordinated direction, which could result in organizational sprawl.

The vision for future agentic infrastructure is an invisible, deeply embedded substrate where agents work proactively and autonomously, with common interfaces rather than distinct tools. Agents would understand team preferences and workflows, becoming responsive to outcomes and budgets rather than explicit instructions.