# The Harness Becomes the Capability Layer

An increasing share of measured agent capability comes from open scaffolding, memory and skill management around frozen model weights, moving competitive advantage away from checkpoints and toward runtimes that anyone can copy.

- Conviction: 37 / 100 (strengthening)
- 7-day move: +11
- Horizon: Emerging (watchlist)
- Tracking since: 2026-08-31T00:00:00.000Z
- Last updated: 2026-09-14T14:04:09.672Z
- Canonical: https://polylog.news/ai/trends/agent-harness-as-capability-layer
- Publisher: Polylog
- Affected regions: Global

## Recent score history

- 2026-09-13: 32
- 2026-09-14: 37

## Recent evidence

- [confirms] A Controlled Test Finds No Advantage for Vendor-Native Coding Harnesses (2026-09-14): A controlled study of paired runs on 80 private tasks put Anthropic's own agent software within 1.25 percentage points of an open-source harness on the same model, with the gap falling in both directions. If the vendor's proprietary scaffolding carries no measurable advantage on its own weights, the scaffolding layer is commoditized and capability attribution shifts back to the checkpoint plus freely copyable runtime conventions.
- [confirms] OpenAI Opens the Codex Harness as a Managed Agents API (2026-09-11): OpenAI released the Codex harness as a managed Agents API in public beta with no platform fee beyond tokens and tools, and lets developers run the agent loop in OpenAI sandboxes or on Cloudflare, DigitalOcean and Oracle infrastructure. Pricing the scaffolding at zero is an admission that the harness is not the revenue layer — OpenAI is commoditizing the runtime to keep token demand attached to its weights.

3 more evidence entries, the full score history, the conviction-driver timeline, and affected assets are for subscribers: https://polylog.news/pricing