The Agent Harness Landscape and the Shift to Markdown-Defined, Event-Driven Agents

Narrative

The financial-press picture is dominated by money and usage figures, not harness architecture. The searches did not return full-text FT, WSJ, Reuters or Economist articles. Only one Bloomberg page (12 March 2026) was retrievable directly. Most other figures arrive second-hand through aggregators that cite Reuters or Bloomberg, so they are leads, not primary evidence. Anthropic's own Series G announcement of February 2026, as relayed by Constellation Research, is the cleanest vendor-reported anchor. It put company run-rate revenue at $14bn, Claude Code run-rate above $2.5bn (doubled since 1 January), and more than 500 customers spending over $1m a year.

Later Anthropic numbers rest on weaker sourcing. Aggregators say Reuters and Bloomberg reported a run rate above $30bn in April, above $47bn in May and above $65bn at the end of July 2026, the last from an investor update. They also say an IPO prospectus showed roughly $4.6bn of 2025 revenue. A Claude Code figure of about $8bn for May 2026 appears only on a statistics blog with no traceable primary source, so treat it as unverified. On OpenAI's side, the cleanest series is vendor-reported weekly Codex users: over 2m (31 March 2026), over 4m (21 April), over 5m (2 June, about 20% knowledge workers). Later 10m, 20m and 25m claims use different or undefined metrics and should not be chained onto that series.

Cursor shows where investors put harness value. Bloomberg reported on 12 March 2026 that Cursor was in talks at about $50bn, nearly double its autumn 2025 valuation (the Series D was $29.3bn in November 2025). Secondary sources say SpaceX took an option in April 2026 to buy it for $60bn, exercised it on 16 June in an all-stock deal, and closed on 14 August. One mirror of Wikipedia attributes the April option to xAI, so the counterparty entity needs a primary filing. The roughly $4bn ARR figure is an estimate from data-aggregator sites, not a disclosure.

Subscription policy shaped harness choice more than features did. From 4 April 2026 Anthropic stopped Pro and Max limits covering OpenClaw and other third-party harnesses, requiring pay-as-you-go "extra usage" (The Next Web, citing The Verge-reported emails). Anthropic's stated rationale was that subscriptions were not designed for these usage patterns. Secondary sources claim $200 Max plans were running $1,000 to $5,000 of agent work, but that is anecdotal. OpenClaw's creator, Peter Steinberger, joined OpenAI in February 2026 and moved OpenClaw to Codex OAuth. This is the clearest evidence that lab pricing is a lever over harness adoption.

On skills, the evidence is mostly security research, not market sizing. An arXiv taxonomy (April 2026) says Agent Skills were adopted within months by Cursor, GitHub Copilot and Gemini CLI, with no mandatory marketplace review and a near-zero authorship barrier. HiddenLayer says skills are not signed and are rarely vetted. A curated security list cites a 98,380-skill analysis finding 157 confirmed-malicious skills, and Trail of Bits scanner bypasses against ClawHub and skills.sh. An earlier study cited by SafeDep found 26.1% of 31,132 analysed skills had at least one vulnerability. No retrieved source in this lane gave verified star counts for anthropics/skills, obra/superpowers or the other seeded repositories, so any such tallies remain untraced here.