Briefing, 22 October 2026: the model matters more than the harness

What changed

• The harness is not where coding intelligence comes from any more. Jediah Katz (SpaceXAI, ex-Cursor) and two independent papers agree: harness choice barely moves pass rates on today's models, but tokens per solved task vary up to 40x. Anthropic and Cursor both published lists of scaffolding they deleted. What the harness still owns: durable cloud runtime, trust and permission scoping, cost, and orchestrating fleets.
• Coordinator plus workers won. Cursor Projects (September) ships a coordinator that never writes code and subscribes to Slack, schedules and PRs. Grokbot and OpenAI's software factory are the same shape. Flat swarms and integrator roles failed.
• MCP went stateless on 2026-07-28. Sessions, the initialize handshake and server-pushed elicitation are gone; Python SDK 2.0 renamed FastMCP to MCPServer with no shim. Anything pinned to mcp<2 is on the maintenance line.
• Compaction and memory moved into the platforms. Both labs compact server-side; memory converged on files with a small index; vendor memory benchmarks miss their claims by 3 to 28 points in reproductions.
• Benchmarks stopped being a signal. OpenAI retired SWE-bench Verified; Epoch rates both SWE-bench variants flawed. Use your own 20 to 50 tasks with pass^k and cost per task.
• For documents, agentic search closed most of the gap to GraphRAG. Microsoft's GraphRAG is in maintenance mode. A cheap PPR graph (HippoRAG 2) beats community summaries at a tenth of the indexing cost. No frontier lab's research product uses a knowledge graph.

Worth reading in full

• Anthropic, "Managed Agents" (2026-04-08) and "Harness design for long-running apps" (2026-03-24).
• Cursor, "Cloud agent lessons" (2026-06-02) and "Projects" (2026-09-10).
• MCP changelog 2026-07-28 and the Python SDK migration guide.
• Fan et al., "Do We Still Need GraphRAG?" (2026-04-01).
• Redis agent-memory benchmark report (2026-09-21), for the reproduction gaps.

AI-assisted briefing: compiled with AI from the week's papers, posts and releases, with links to each source.