Your commit costs $12.69 and your harness engineer knows why

Cursor published a number this week that should make every “just pick the best model” person uncomfortable: a single commit through Fable 5 costs $12.69. The same commit routed through Cursor Router, their new model dispatcher trained on 600K+ live coding sessions, costs $6.76. Same code, same repo, nearly half the price. The difference is entirely in the harness.

The metric nobody was tracking

Cursor also introduced a concept they call keep rate: what percentage of agent-generated code actually survives in the codebase over time. A user moving on to the next feature is a positive signal. A user pasting a stack trace is a negative one. They’ve been using keep rate alongside user satisfaction classifiers to A/B test every harness change and every model launch for the past nine months.

This matters because it kills the vibes. You can’t argue with code that got deleted. And Cursor is measuring it across millions of requests per week, per model, per tool, per harness variant. They drove tool-call error rates down to 2-3 nines in a focused sprint. They found one model developing “context anxiety,” refusing to work as its context window filled up, and fixed it with prompt adjustments. These are SRE problems, solved with SRE discipline.

GitHub landed on the same conclusion from a different direction. Burke Holland’s harness workflow boils down to eight steps, and the most interesting one is the rubber-duck review: after implementation, Copilot requests a review from a different model family. GPT wrote it, Sonnet reviews it. Different training data, different blind spots. The dual-model pattern transfers to any agent tool, and it lives entirely in the harness.

The tax you’re paying in silence

Meanwhile, your agent is burning tokens on tools it never calls. Every tool schema rides along in every request. A 200-tool catalog at 250 tokens per schema across a twelve-turn task spends 600,000 input tokens on tool definitions alone. The actual conversation? Maybe 20,000 tokens. Microsoft Foundry shipped Tool Search to fix this: a retrieval step that loads only the tools an agent needs per turn. Cursor’s version is dynamic tool calling, where most tool descriptions get lazy-loaded the first time the model reaches for them.

Both are pure harness optimizations. The model stays the same.

Harness engineering gets a name

Thoughtworks’ Kief Morris gave a talk at PlatformCon calling this exactly what it is: harness engineering. His framing: build CI/CD pipelines that enforce quality automatically (“on the loop”) so humans supervise the system instead of rubber-stamping every diff. He also coined two terms worth remembering: cognitive debt (the team’s understanding gap) and intent debt (the rationale nobody wrote down). Both accumulate faster when agents are shipping code at machine speed.

The pattern is converging from every direction. Cursor tunes per-model prompts so aggressively that dropping reasoning traces from Codex causes a 30% performance hit. GitHub stacks agent sessions into chained PRs for legacy migration. Temporal 5x’d their AI spend and their CEO openly admits he can’t prove it caused their revenue to double.

Everyone is spending more on agents. The ones measuring cost per commit already know whether it’s working. The rest are vibing.

Emerging

• MCP 2.0 kills session state. The release candidate removes initialization handshakes, session IDs, and three deprecated features, letting remote MCP servers deploy as ordinary stateless HTTP services behind round-robin load balancers.

• Ontologies resurface as agent guardrails. A UC Berkeley professor argued at AI Engineer World’s Fair that OWL and Schema.org give probabilistic agents the structured boundaries they keep failing to enforce on their own, with Neo4j’s CEO echoing the same “thin agents on a shared semantic layer” architecture.

• PortSwigger cages its pentesting agents. Burp AT ships agentic AI for professional penetration testing behind a deterministic control layer where agents propose actions but a separate enforcement system governs scope, tool access, and approval rules.

Corporate stuff

• GPT-5.6 prices crater, again. Luna dropped 80%, Terra 20%, and GPT-5.6 Sol now rewrites its own inference kernels to cut serving costs. March’s flagship intelligence costs roughly 1/13th the token price four months later.

• Codeberg votes to ban vibe-coded projects. The open-source Git host’s member assembly passed a motion prohibiting repositories “mostly consisting of LLM-generated code,” citing resource strain and copyright concerns. Theo’s breakdown lands the strongest counterpoint: the CI/CD cost argument is real, everything else is politics.

• Kimi K3 opens weights few can run. Moonshot released a 2.8-trillion-parameter MoE model on Hugging Face that matches Fable 5 on coding benchmarks at roughly a third the API cost. Self-hosting requires eight or more H100 servers, so “open weights” here means downloadable if you have a data center.

• India pricing race heats up. Cursor launched a $7/month India plan with UPI payments and its own Composer model (no frontier model licensing costs), OpenAI already had ChatGPT Go at $4, and Anthropic showed up with rupee billing that somehow costs more than the US price.

• Replit expands from code to design. Replit Design launched with what the team calls Ambient Intelligence, wrapping Claude, GPT-5, Gemini, Kimi, and GLM around a Mobbin-powered UI reference library of 600,000+ real-world screens. The play is clear: own the full path from idea to published app so no one ever leaves.

Underground

• Willison builds three MCP tools in a week. Stateless MCP re-energized Simon Willison enough to ship mcp-explorer, datasette-mcp, and llm-mcp-client in rapid succession, each small enough to understand in an afternoon. His reasoning: MCP tools are easier to audit and control than giving agents a shell and curl.

• Zero-knowledge proof of human typing. A developer built PoHI, a zk-SNARK protocol that proves text was physically typed by a human based on keystroke rhythm, without revealing the keystrokes themselves. The threat model doc openly admits where it breaks and invites adversarial testing.

• Final Fantasy meets agent monitoring. A Canadian realtor used AI to build an RPG-styled dashboard for tracking AI agents, where “HP” means headroom points (remaining context window capacity). Plain HTML, no build step, MIT licensed.

• Agent finds the cloud waste, nearly deletes the backups. A practitioner gave an agent read-only AWS/GCP access and one prompt: cut the bill 20%. It found 14% in obvious waste within an hour, then confidently proposed deleting the disaster-recovery snapshots.