A neon-lit studio with a many-headed Squad HydraFusion machine, a developer at the desk, and two dogs curled up nearby.

Squad 1.1.0 isn’t out yet, but it’s close, and I wanted to tell you what’s in it before it lands. The headline is something a lot of people in the Squad community have been asking for since HydraFusion showed up: Squad is getting HydraFusion support, opt-in, for the work where it matters most. The release also changes how Squad picks models for every agent on your team, and I think that part will end up mattering to more people than the HydraFusion bit. So this post covers both.

A word about the version number. We shipped 1.0.1 a couple of weeks ago, and I said then that stable software still improves one careful patch at a time. This one isn’t a patch. It changes a default: Squad will start picking models quality-first instead of cost-first, and that can change what you spend without you touching anything. When we change a default on you, that deserves more than a patch number, so we’re calling this 1.1.0. All of it is in #2225, which closes #1720, #1598, and #2014.

why this happened

When HydraFusion landed in VS Code and the GitHub Copilot app, folks in our community channel tried it right away. It’s a Copilot model that orchestrates several models inside one turn, and people liked what it did. Then they hit a wall. One of them put it plainly in early September: “for now, we have to choose between Squad or Hydra … I wonder how is everything going to fit together.” A month later, the same person was still stuck: using both at once didn’t work well, and “Hydrafusion is a squad on its own.”

That stung a little, mostly because it was right. Squad was picking models per agent with a set of rules that had grown up over time in a few different places, and those rules leaned cost-first. If you picked a great model for your session, a sub-agent could still end up on something cheaper. If you picked HydraFusion, Squad didn’t really know what to do with it. I said in the channel that if Squad wasn’t choosing the models people thought were best, we should change the behavior. 1.1.0 is that change.

one rulebook for every agent

The foundation of 1.1.0 is a single, central model resolver. Every agent spawn will run through the same precedence, and the first match wins:

  1. User override — “use opus for this task.” One spawn, nothing persisted.
  2. Persistent config — “always use opus,” saved to .squad/config.json, including per-agent overrides.
  3. Session model — the model you picked for the session. Every sub-agent inherits it.
  4. Charter preference — the ## Model section in an agent’s charter.
  5. Task-aware auto-selection — Squad looks at what the agent is doing and picks a tier.
  6. Default — the head of the standard chain.

The third one is the fix people felt most. If you pick a model for your session, the agents Squad spawns for you will use it instead of quietly dropping down a tier. And every spawn will tell you what it picked, where that choice came from, and why:

📋 Scribe (claude-haiku-5.5 · task-auto) — logging session — mechanical task → fast tier

That line is small, but I like it a lot. You never have to guess which model an agent used or why. Fallback got stricter too. If a model isn’t available to you (plan, org policy, rate limit), Squad moves forward through its chain and never back toward the head, never up a tier, and never past your cost ceiling. If nothing’s left, Squad leaves the model off and lets Copilot choose.

quality first is now the default

This is the change I want you to know about before you upgrade, and it’s the reason this is a minor release and not a patch. Squad used to lean cost-first. In 1.1.0 the default profile is quality-first: architecture, research, verification, review, security, and design work go to premium models; code, docs, and planning go to standard; and only mechanical work like Scribe, git, and changelogs stays on the fast tier. The reasoning is simple. An underpowered model on an architecture review doesn’t save you money if you end up running it three times.

If you’d rather keep spend down, you still can. balanced and economy are one sentence away (“use the balanced profile” or “switch to economy mode”), and economyMode: true in your config still works. Profiles only change what Squad picks automatically. They never replace a model you chose yourself.

We also refreshed the catalog. Squad now recommends only public GA models from the GitHub Copilot supported models table, so the newer Claude, GPT, Gemini, Grok, and MAI models are in, and retired names like claude-sonnet-4.5 and gemini-3.1-pro are out of the recommendations. If you name an older model explicitly, Squad still passes it through.

and now, HydraFusion

Here’s the part the community asked for. In 1.1.0 you’ll be able to turn on HydraFusion with a new profile:

{
  "version": 1,
  "modelProfile": "hydrafusion"
}

…or just for the session: “use the hydrafusion profile for this session.”

With that profile on, the six kinds of work where quality matters most (architecture, research, verification, review, security, and design) go to hydrafusion first, with the quality-first premium chain behind it. Everything else follows quality-first as usual. Scribe stays on a fast model, because nobody needs a committee of frontier models to append a line to a log. Tamir shared a kid-friendly explanation in the channel that started with a picture I like: Squad’s agents are robots with a box of plug-in brains. You pick the brain for the session, and the sub-agents use it too. Turn on HydraFusion, and the hardest jobs bring multiple models together to debate and combine “like the Power Rangers forming a Megazord.” Coding and Scribe stay on fast brains so routine work keeps moving.

You can also skip the profile and pick hydrafusion directly, as your session model, as a defaultModel, for one agent, or for one task. Squad honors it like any explicit choice, and a session pick flows down to every sub-agent.

A few guardrails, all on purpose:

  • It’s opt-in. HydraFusion is a research preview, so it isn’t in any default chain or GA profile. Squad will never escalate you into it on its own.
  • Fallback tells you. If HydraFusion isn’t available to you, Squad falls forward to the next GA premium model (claude-opus-5.5) within your cost ceiling, tells you once, and doesn’t keep retrying.
  • Visual work isn’t routed to it automatically, since its vision support isn’t documented yet.
  • VS Code runs one model per session. Per-agent routing applies in the Copilot CLI and the Copilot app. In VS Code, pick HydraFusion in the model picker if you want it everywhere.

Since it’s a preview, check the GitHub changelog for current requirements. Business and Enterprise orgs need an admin to turn on preview features, and your client may need HydraFusion enabled too.

how we got here

This one was a group effort, and the way it came together was a pretty good picture of how we build Squad now. I drafted the design, and Tamir helped make the changes easier to explain. His robot-brains analogy captures the session-model inheritance in a way the resolver diagram never will. His Megazord comparison is a pretty good description of what HydraFusion does on the hardest assignments: several models debate and combine their work.

The reviews weren’t just vibes, either. The cost change from the new default came up as something users should hear about up front (hence the section above), and the questions that followed (should fallback be silent? which roles get HydraFusion? what about VS Code?) helped shape the behavior in the PR. We’re finishing the last review items and hands-on testing in the Copilot app, CLI, and VS Code before we cut the release.

Thank you to everyone in the Squad community who kept pushing on this with real feedback, especially the folks who kept trying both tools and telling us honestly where it fell apart. Thank you to Tamir, as always, for the review and for the Megazord. You asked for it, and I’m grateful you kept asking.

when it ships

Once 1.1.0 is out, upgrade and tell your coordinator what you want:

use the hydrafusion profile for this session
always use the balanced profile
what model did that last agent use, and why?

Until then, the full details are in #2225 and its updated model selection docs. If you want to kick the tires early, review comments on the PR are welcome. After it ships, if something picks a model you didn’t expect, open an issue. The rationale line in the spawn output should make it easy to tell us exactly what happened.