US government cleared Anthropic's Mythos 5 for ~100 companies and federal agencies (Fri Jun 26); OpenAI shipped GPT-5.6 the same day to ~20 government-approved partners. The best closed models now arrive on Washington's schedule.
Why governments got nervous (context, June 2-3): Project Glasswing partners surfaced 10,000+ high/critical vulns since early April; Mozilla found and fixed 271 Firefox vulns in testing (>10x a prior version); Cloudflare found 2,000 bugs incl. 400 high/critical.
Operator takeaway: leverage is not frontier access (you may not get it). Leverage is owning the workflow on whatever model you can reach. That is exactly where today's lead lands.
The stories
Dynamic workflows go DEFAULT in Claude Code (Anthropic)
The ultracode setting sends xhigh reasoning + auto-orchestrates dynamic workflows: Claude writes a JS script, a runtime fans out tens-to-hundreds of parallel subagents (up to 16 concurrent, 1,000/run). We covered the May 28 launch; the DELTA: it's now ON by default for Max/Team/Enterprise + the API (Pro: enable in /config).
Workload Identity Federation on the Claude Platform (Anthropic, this week)
Replaces static API keys with short-lived, scoped credentials issued at request time — no static creds to create, rotate, or leak. Workloads auth with the identity they already have (AWS IAM, GCP/K8s service account, Azure managed identity, GitHub Actions, Okta, OIDC). Each workload gets its own service account, roles, and audit trail.
The move: wire it into one CI pipeline, delete the static key.
OpenAI's GPT-5.6 (Sol flagship / Terra ~GPT-5.5 at 2x cheaper / Luna cheapest) is preview-only to ~20 govt-approved partners; broad release "coming weeks." Practical outcome: GPT-5.5 remains the OpenAI production API anchor; pinned code won't auto-migrate. Build assuming the model you ship on may not be the best model that exists.
Run one scoped dynamic workflow with the `ultracode` keyword, watch the dashboard to learn the cost, save the script if it's good. You walk away with a reusable command and a real feel for crew-of-agents pricing.
The written brief from each episode, free, every morning.
Transcript
Nova: Claude Code can now decide, on its own, to fan a single request out across hundreds of parallel agents, and on the Max, Team, and API plans that behavior is now on by default. The thing that ported seven hundred fifty thousand lines of Bun from Zig to Rust is now the default mode, not a hidden flag. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Sunday, June twenty-eighth, twenty twenty-six. I am Nova.
Dex: And I am Dex.
Nova: On today's Mainframe: dynamic workflows go default in Claude Code, and the violet keyword that triggers them, ultracode, can drain your whole token budget in ten minutes if you are not careful. Plus, Anthropic quietly shipped a way to kill static API keys entirely.
Dex: Then the velvet rope tightens: Mythos restored, GPT five point six staggered, and what that means for the model you actually ship on. And two tools you can drop in today, including a Google spec that makes your agent stop guessing what your app should look like.
Nova: Let's get into it.
Dex: One honest note before we go deep: our voices are AI. The reporting, the picks, and every opinion here are human-made. Now, the board.
Nova: Here is the ground today. The frontier is being rationed. On Friday, the US government granted Anthropic permission to release its Mythos 5 model to a group of roughly one hundred companies and federal agencies. Same day, OpenAI shipped GPT five point six, and limited access at the behest of the US government to around twenty companies, whose participation was approved by the government. That is the macro: the best closed models now arrive on Washington's schedule.
Dex: And the defense layer that triggered all this is real. Earlier in June, Anthropic expanded its cyber program, and the numbers are why governments got nervous. Claude Mythos Preview surfaced more than ten thousand high or critical software vulnerabilities since the program launched in early April. Mozilla found and fixed two hundred seventy-one vulnerabilities in Firefox while testing the model, more than ten times the number found in a previous version.
Nova: So why does the builder care? Because the engine that finds ten thousand bugs is the same engine now sitting in your terminal by default. The leverage is not access to the frontier model, which you may not get. The leverage is owning the workflow that runs on whatever model you can reach. That is exactly where today's lead story lands. Anthropic just made multi-agent orchestration the floor, not the ceiling, of what Claude Code does.
Dex: Which brings us to the spot.
Nova: Today's Mainframe is brought to you by Outpace. Everything on the board today is consolidation: the frontier rationed, the giants merging, access gated. Outpace is the other direction. Picture an operator who spent thirty years building real businesses: real estate, hospitality, food, even a theater. Now he aims a full team of AI agents at one client's project at a time, strategy to ship. Outpace is a system, not a freelancer and not a faceless mega-agency: the agents do the work, one senior operator directs them, and at the end you own the code and the keys. If the rationing on the board today makes you want a partner who builds you something you control, book a thirty-minute call at outpace dot media. That is outpace dot media.
Dex: The Wire. Lead story, and it is the one you can act on today. Anthropic shipped dynamic workflows into general use in Claude Code, driven by a setting called ultracode. It is accessible through the effort menu, and it sets the effort level to extra-high while letting Claude decide automatically when to use a workflow to handle your task.
Nova: Decode that. Normally one agent works your request in one conversation, one task at a time. Dynamic workflows let Claude write a JavaScript script and have a runtime execute it in the background, fanning out tens to hundreds of parallel subagents. Think of it as Claude stops being one carpenter and becomes a foreman who hires a crew, gives each worker one room, then has two other workers inspect every room before it is signed off. Jarred Sumner used dynamic workflows to port Bun from Zig to Rust: roughly seven hundred fifty thousand lines, eleven days, ninety-nine point eight percent of the existing test suite passing. That is the upper bound of what has been demonstrated publicly.
Dex: We covered the original launch back in May. What is new today, and the reason it leads: the defaults flipped. Dynamic workflows are now on by default for Max, Team, and Enterprise plans, and when using Claude Code via the API. Pro plan users can enable them in config. So the question is no longer whether you opt in, it is whether you know when it fires.
Nova: And here is the catch, because this is a builder's brief, not a hype reel. The cost is brutal if you are sloppy. One developer on the two-hundred-dollar Max plan reported burning twenty percent of their entire weekly limit on day one, just from turning on dynamic workflows. One Pro user on the twenty-dollar plan reported hitting their cap after about ten minutes. The fix is discipline. Reserve ultracode for work you would otherwise hand to several engineers in parallel. If it is one person, one afternoon, ultracode is overkill and a single agent at high or extra-high is the right tool.
Dex: So the move this week: do not flip ultracode on for your whole session. Use the per-task trigger. Include the keyword ultracode in your prompt to run a single task as a workflow without changing the session's effort level. Point it at one genuinely big job: a codebase-wide auth audit scoped to one directory. And if it fires when you did not mean it, press Option-W on macOS or Alt-W on Windows and Linux to dismiss it.
Nova: There is a second, quieter Anthropic ship this week that agency operators should not miss. Workload Identity Federation on the Claude Platform. It replaces static API keys with short-lived, scoped credentials issued at request time. With it, there are no static Anthropic credentials to create, rotate, or leak.
Dex: Translate that for the keyboard.
Nova: A static API key is a house key you mail to a contractor and hope they never copy. Workloads authenticate with the identity they already have: an AWS role, a GCP or Kubernetes service account, an Azure managed identity, a GitHub Actions token, Okta, or other standard providers. Each workload gets its own identity, roles, and audit trail instead of a shared key. If you are running Claude inside client infrastructure, this is the difference between hoping a leaked key does not surface and never having one to leak. Wire it into one CI pipeline this week and delete the static key.
Dex: Quick hit on the velvet rope, since it moved again. OpenAI says Sol is its strongest model yet, Terra is for everyday use with similar performance to GPT five point five at twice as cheap, and Luna is its lowest-cost variant. But you cannot get them. The practical outcome is that GPT five point five remains the current OpenAI production API anchor; any code pinned to it will not auto-migrate until OpenAI promotes the new line. Build assuming the model you ship on may not be the best model that exists.
Nova: Tool Lab. Two picks, both you can run today. First, the trending one: DESIGN dot M D, from Google Labs. It is a format specification for describing a visual identity to coding agents, giving them a persistent, structured understanding of a design system. It sat at roughly nineteen thousand stars this week.
Dex: Why it matters. Every time your agent builds a screen, it guesses your brand from scratch: wrong blue, wrong spacing, wrong font. A DESIGN dot M D file combines machine-readable design tokens in YAML front matter with human-readable rationale in markdown prose. Tokens give agents exact values; prose tells them why those values exist and how to apply them.
Nova: The move: drop a DESIGN dot M D in your repo root with your real colors, type, and spacing, then tell Claude to build to it. And there is a validator. It validates a DESIGN dot M D against the spec, catches broken token references, checks WCAG contrast ratios, and surfaces findings as structured JSON that agents can act on. One command, run it with N-P-X at google slash design dot M D. Credit to the Google Labs Code team.
Dex: And the hidden gem, not on the front page. There is a community repo, awesome-design dot M D by VoltAgent, that ships ready-made DESIGN dot M D files reverse-engineered from real brands. Copy one into your project, tell your AI agent build me a page that looks like this, and generate high-quality UI that stays visually consistent, built with analyzed patterns, tokens, and rules, not surface-level outputs. Stripe, with its signature purple gradients, is in there, alongside Airbnb, Linear-style tools, dozens more. So if a client says make it feel like Stripe, you hand the agent the Stripe DESIGN file instead of describing it in prose and praying. Credit to the VoltAgent team.
Nova: The Deep Dive. The through-line of this whole week is one shift: AI stopped being a single assistant you chat with and became a crew you direct. Dynamic workflows are the clearest proof. Opus broke the one-model, one-conversation, one-task pattern on May twenty-eighth, and agents went from assistants to something closer to architects. But the reason this is now default, and the reason Washington is rationing the frontier, is the same fact: a crew of agents is powerful enough to be dangerous. The engine that ports Bun in eleven days is the engine that finds ten thousand zero-days.
Dex: So how does a builder actually ride this instead of getting flattened by it?
Nova: Concretely, this week. Step one: turn on dynamic workflows deliberately on one real job, not your whole session. Pick a task that genuinely decomposes: audit every endpoint under one directory for missing auth checks. Scope it to that path so the token bill stays sane. Step two: watch the dashboard. Open the workflows dashboard and every phase appears with the agent count, the total tokens consumed, and the elapsed time. Learn what a real run costs before you point it at anything big. Step three: if the run finds a good shape, save it. Press S in the workflows dashboard to save the generated script to your workflows folder and invoke it later as a slash command. That is the compounding move: today's audit becomes tomorrow's reusable command.
Dex: And step four, the unglamorous one: lock the keys before you let a crew of sixteen agents touch client code. Wire Workload Identity Federation into the pipeline so there is no static credential sitting in the blast radius. Crew plus scoped, short-lived identity. That is the safe version of this power.
Nova: Your level up this week: run one scoped dynamic workflow in Claude Code with the ultracode keyword, watch the dashboard to learn the cost, and save the script if it is good. You will walk away with a reusable command and a real feel for what a crew of agents costs.
Dex: And if you want every story, number, and link from today in your inbox, the free Mainframe newsletter drops the full written brief each morning. Subscribe, it is the written version of everything we just ran.
Nova: That is today's Mainframe. We will see you tomorrow.