3 AI Harness Categories Exist—Only One Doesn't Trap You
✦ Report Highlights
The industry frames AI coding harnesses as a choice between speed (general-purpose) and control (custom-built). That framing is wrong. The real distinction is closed versus composable and transparent. There are three categories of harness — locked-core general-purpose, fully custom, and composable substrate — and only the third avoids both the forced fork and the framework-team trap. Agent Zero has been building this third category since mid-2024, shipping orchestration, skills, memory, isolation, and supervisory GUI before the industry converged on the same ideas. The question teams should ask is not "build or buy" but "does your harness…
Forked roads fade at dawnClear parts compose what lastsChoice becomes the cage
▶ About this episode
The AI coding harness debate of 2026 has settled into a familiar binary: speed versus control. Build or buy. General-purpose or custom. This video explains why that framing is fundamentally wrong—and reveals the third category most teams haven't even considered.
Show more
Based on a strategic brief unpacking Agent Zero's "composable substrate" thesis, this breakdown walks through the three actual categories of AI harness architecture: locked-core general-purpose tools that force a fork the moment you need something unexposed, fully custom builds that quietly turn your product team into a framework team, and the composable substrate approach that lets upstream improvements flow under your overrides without forcing a fork.
We cover the five questions every team should run against their current harness, the four surfaces that define a truly "general-purpose" system (rules, skills, tools, memory), and why the real distinction isn't general versus custom—it's closed versus composable and transparent. We also examine how Claude Code, Cursor, and Codex have converged on ideas Agent Zero shipped in mid-2024, and what that convergence means for teams making harness decisions today.
Key takeaways
- The speed-versus-control binary is a structural limitation, not a feature choice
- Plugins and rules are not ownership—whoever owns the orchestration core owns your roadmap
- Isolation, trust mechanisms, and reproducibility must be architectural, not bolted on
- The question isn't "build or buy"—it's "does your harness force you to choose?"
If you're evaluating AI coding harnesses in 2026, this reframing changes the decision entirely.
