The coding agent became a commodity, now the terms are the pitch

Meta priced the data clause, and now yours has a number

Between Wednesday and Friday, three terminal coding agents launched: Meta’s Muse Code (opens in a new tab) in beta, Memcode in public beta, and Clark Code as Apache-2.0 open source. Read the three pitches side by side and notice what none of them says: ours writes better code. Meta competes on price, Memcode on persistent repository memory, Clark on openness. A fourth launch in the same window, Mirafold, does not sell an agent at all. It sells a browser UI over the incumbents, and an accessory market is the oldest sign there is that a category is commoditising underneath. The pattern item has the receipts.

The sharpest move is Muse Code’s pricing table. Standard pay-as-you-go is $1.25 per million input tokens and $4.25 per million output. A “contributor” tier prices the same product at $0.10 and $0.20 (opens in a new tab), roughly 12x cheaper on input and 21x on output, in exchange for Meta training on your prompts and completions. Meta AI chief Alexandr Wang pitched cost as the sell against Claude Code and Codex, but the structure is the story: as far as we can find, this is the first coding-agent vendor to price a discount explicitly as a data trade rather than as volume or seats.

That makes a training-data policy a priced variable for the first time. Until Wednesday, “we don’t train on your code” was a compliance line buried in a trust center. Meta just published its market rate: about 90 to 95% off list.

Your move: if you sell an AI devtool, decide your answer to data-for-discount before a competitor prices it for you, and move your training-data position out of the FAQ and onto the pricing page, where it now does positioning work. If your pitch still leads with capability, this is the week to re-read it: when features converge, the terms are what is left to compare.

The platform’s gate and its exit are now public numbers

Two smaller stories are the same story: the terms of being on a platform are becoming comparable numbers.

At the front gate, developer Josh Symonds published a dated timeline of his MCP connector’s path through Anthropic’s directory review: submitted March 22, told it was “in the review queue” as late as June 15, then told July 30 that the queue had been discarded and he should resubmit, while OpenAI approved the same connector in 29 days. One account, not a pattern, but the first concrete stranded-integrator story since the August 4 piece argued the directory is a gated distribution channel, and it puts numbers on the gate.

At the back door, GitHub gave Spark users 27 days’ notice: signups frozen immediately, export by August 31, and apps built on the llm() function already broken because the service behind it retired July 30. The shortest window yet in the ledger this desk keeps (MCP: 12 months; GitHub Models: six weeks; HCP Vagrant, announced the same week: about 10 months). Wednesday’s dive made the argument in full: the window is the headline, but the mechanics decide who gets burned.

Your move: if a platform directory is in your channel mix, its review queue is now a risk line item with a date on it. Dual-list where you can, and keep an install path you own so no launch waits on someone else’s queue.

Publish the number that hurts

The content move of the week is Sentry’s build-in-public post on routing its Seer agent’s bug-fix PRs to the right engineer through Slack. The numbers are modest and useful, action rate on those PRs up about 21%, 48-hour response rate up about 13%, but the reason the post travels is the third number: close-without-merge rose 12.5%, and Sentry printed it, then argued why that is healthy (engineers often chose a broader fix than the agent’s narrow one). Seer costs $40 per active contributor per month, so this is commercial content. It earns the read anyway, because the uncomfortable number makes the flattering ones legible. The dogfooding-receipt pattern again: proof you can check beats a claim you must trust.

Your move: next time you publish results, find the number that cuts against you and print it beside the ones that flatter, with your reading of why. A claim that costs you something reads as proof.

Quick hits

• Vercel had a platform week: Agent Plugins 1.0.0, a vendor-neutral packaging format refined with AWS, Anysphere, GitHub, Microsoft and OpenAI, package a skill once instead of per client; plus v0 as an API and AI Gateway on AWS Marketplace, the last aimed squarely at procurement as the real adoption blocker. Watch whether the non-signers (Claude Code, Zed, JetBrains) adopt the format.
• Okta’s Cross App Access protocol, with 25+ launch partners, is live natively in Auth0 tenants: a bid to own the auth layer under enterprise MCP deployments before the spec’s CIMD migration forces the question.
• Node.js shipped its first API-docs redesign with search, offline support and llms.txt. The machine-readable-docs default reaching the most conservative surface in the ecosystem.
• The Emacs project drafted an AGENTS.md that says don’t generate, instructing LLM assistants to search and analyse rather than write code, and to surface its no-LLM-contributions policy. The first documented case in our roster of AGENTS.md used to restrict agents rather than enable them.
• The DevRel hiring-bar thread retired on schedule. In W29 one practitioner claimed DevRel postings now demand measurable ROI, with corroboration due from DevRelCon NYC recaps by mid-August. Eighteen days after the conference wrapped, no recaps have published at all. One person’s job search.
• Shelf housekeeping. If you installed devmarketing-skills from our skills shelf, the publisher changed its install line to npx add-skill jonathimer/devmarketing-skills (was npx skills add …). Both cards are fixed; the repo itself is alive and unchanged.

Watch

1. Someone answers the data trade by end of August, or nobody does. The pattern item’s call: an incumbent, Anthropic, OpenAI or GitHub, ships or publicly rules out a data-for-discount tier. A match confirms the squeeze; explicit refusal makes “we don’t train on your code” the counter-positioning at full price. Silence past September says the entrants are noise.
2. Cerebras’ free-tier sunset lands August 17. The test, set in the deprecation dive: a clean landing returns deprecation-named errors against a stable model catalog; a dirty one is silent 404s. Checkable the same day.
3. The MCP directory gets a second account or an SLA by end of August. Either another integrator publishes a dated timeline corroborating the Symonds account, or Anthropic publishes review-queue expectations, an SLA, an appeals path, a status page. A second account makes it a pattern; a published SLA closes the story in Anthropic’s favour.