Prediction Ledger

The official MCP registry (registry.modelcontextprotocol.io) will list more than 10,000 distinct servers by 2027-01-03.

Resolves: The official MCP registry's public API or web UI reports a total server count above 10,000 on or before the resolve date; a dated archive capture or API response screenshot suffices.

GitHub's 2026 Octoverse report will state that a majority (over 50%) of surveyed or measured developers' code output involved AI assistance.

Resolves: The published Octoverse 2026 report (github.blog or octoverse.github.com) contains a headline or body claim that >50% of code written/pushed by developers in its data or survey involved AI assistance; exact framing may vary but the majority claim must be explicit.

By 2027-01-03, at least two of Anthropic, OpenAI, or Google will publicly sell an individual subscription tier priced at $300/month or more that is marketed primarily around autonomous or long-running agent usage.

Resolves: Public pricing pages (verifiable via Wayback Machine captures dated on or before 2027-01-03) show two such companies each offering an individual-plan tier at >=$300/month whose marketing copy centers agents or autonomous task execution.

By 2027-01-03, at least two frontier labs will report scores on the same third-party agentic terminal or harness benchmark (such as Terminal-Bench) in their official model announcement posts.

Resolves: Two distinct frontier labs' official model release blog posts, published on or before the resolve date, each cite a score on one identical third-party agentic/terminal benchmark; the posts are publicly readable.

By 2027-07-03, the MCP specification will include a MUST-level (mandatory) requirement specifically addressing tool-description injection or tool poisoning.

Resolves: The MCP specification at modelcontextprotocol.io (or its changelog/commit history) contains normative RFC-2119 MUST language whose stated purpose is mitigating tool-description injection or tool poisoning, merged on or before the resolve date.

By 2027-07-03, at least three major coding harnesses (from: Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Windsurf, Devin) will offer documented parallel multi-agent orchestration as a generally available feature.

Resolves: Official public documentation for three of the listed products, live or archived on or before the resolve date, describes spawning multiple concurrent agents/subagents on one task as a GA (non-beta, non-waitlist) feature.

By 2028-07-03, a company with more than 1,000 employees will publish a public incident postmortem that attributes a production outage to a change authored by an autonomous coding agent.

Resolves: A public postmortem, incident report, or official engineering blog post from a qualifying company explicitly identifies an AI coding agent as the author of the change that caused a production outage; press coverage quoting the company's own attribution also counts.

By 2028-07-03, Anthropic or OpenAI will be publicly reported, via company statement or credible financial press citing company figures, to exceed a $20B annualized revenue run-rate.

Resolves: A company announcement, investor letter, or reporting from an established financial outlet (citing company-provided figures) states annualized revenue or run-rate above $20B for either company on or before the resolve date.

By 2028-07-03, a peer-reviewed paper or major-lab technical report will demonstrate a harness/scaffold change producing a 10+ percentage-point gain on a standard agentic coding benchmark while holding the model fixed.

Resolves: A publicly available paper (arXiv or peer-reviewed venue) or official lab report shows, in its own tables, the same model scoring >=10 points higher on SWE-bench Verified, Terminal-Bench, or an equivalent standard benchmark under a modified harness/scaffold.

By 2028-07-03, at least three venture-funded companies whose primary product is MCP or agent-tool security scanning will each have publicly announced funding rounds of $10M or more.

Resolves: Public funding announcements (company posts, Crunchbase, or press) show three distinct companies, each with a primary product described as MCP/agent-tool security or trust scanning, each with a disclosed round >=$10M announced on or before the resolve date.

By 2028-07-03, an open-weights model that runs on a single consumer GPU with 32GB or less of VRAM will exceed 70% on SWE-bench Verified.

Resolves: A model with publicly downloadable weights appears at >70% on the SWE-bench Verified leaderboard, and its model card or a reproducible community report documents inference (possibly quantized) within 32GB VRAM on a single consumer card.

By 2028-07-03, a publicly listed model will score 90% or higher on the SWE-bench Verified leaderboard.

Resolves: The official SWE-bench leaderboard (swebench.com or its successor) shows any model entry at >=90.0% on the Verified split on or before the resolve date; archived leaderboard snapshots count.

By 2028-07-03, a software company with 5 or fewer employees will publicly document $10M or more in ARR for a product built primarily with AI coding agents, with coverage in at least two major tech outlets.

Resolves: Public statements from the company (blog, interview, or verified social post) claim >=$10M ARR and <=5 employees and primary AI-agent authorship, corroborated by coverage in two established tech publications on or before the resolve date.