harrison chase

Harrison argues that enterprises must own all three components of agent intelligence — model, context, and harness — with the harness as the critical orchestration layer. He presents a spectrum from off-the-shelf harnesses (Claude Code, Codex, Deep Agents) to custom cognitive architectures, driven by out-of-distribution tasks and predictability needs (e.g., financial services).

Private benchmarks (Harbor), trajectory observability (LangSmith), and automated trace curation (LangSmith Engine) form a closed loop where production data continuously improves the harness, model, or context — turning AI deployments into hill-climbing systems that compound value over time.

All AI agents share a core LLM-loop architecture, but competitive differentiation comes from customizing the harness via middleware (hooks, plugins) for domain-specific needs like memory, sub-agents, and file-system access. The more out-of-distribution the task, the more custom harness engineering is required.

Enterprises gain trust not through exhaustive upfront testing but by shipping fast with small eval sets (5-10 cases), observing real runs via tools like LangSmith, and iterating weekly—because agent architectures become obsolete every 9 months.

Enterprises gain trust through observability (LangSmith) and eval-driven development; starting with just 5-10 scenarios forces product thinking about desired agent behavior, and living eval datasets capture real-world usage to guardrail future prompt changes—shipping iteratively with limited blast radius beats months-long waterfall builds.

Agents that run persistently in the background, listen to events (emails, triggers), and spin up long-running sub-agents asynchronously will replace manual copy-paste workflows, delivering step-change productivity gains in enterprises where events fire constantly.

Harrison Chase predicts asynchronous sub-agents managed by an orchestrator will become the dominant coding agent paradigm within months, while always-on agents listening to enterprise event streams (email, Slack, etc.) will deliver step-change productivity by eliminating copy-paste workflows.

The agent stack separates into model, harness (Deep Agents, LangGraph), and runtime (OpenShell, GPU cloud, local); picking best-of-breed per layer and swapping harnesses every 9-12 months as capabilities leapfrog is necessary—architectures from 18 months ago are already obsolete for complex tasks.