Zed 1.22: A Different Brain for Every Subagent — Multi-Model Orchestration Moves Into the Editor

From "pick one model" to "a model per task"

Over the past year, AI coding tools competed on "supporting more models": who integrates the most, who updates fastest. Zed 1.22 (stable, September 30) pushes the question one step further: when you have a dozen models at hand, the real productivity problem becomes how to assign a different brain to each subtask.

The headline change looks small: spawn_agent gains an optional model parameter. When spawning a subagent you can now specify its model directly instead of inheriting the parent agent's model or the global default, with the active model shown on the agent card. The PR (#64498, from community contributor itsfuad) effectively promotes "multi-model orchestration" from a hand-rolled prompt trick to a first-class editor capability.

Why does this deserve its own story? Because it matches the daily reality of vibe coders: the main agent handles architecture and planning while spawned subagents generate code, write tests, review docs — tasks with very different demands on model capability. Code generation may want the strongest reasoning model; churning through test cases is fine on a cheap fast one. Previously you either used the expensive model for everything (burning money), the cheap one for everything (discount quality), or switched manually (exhausting). Now division of labor and model selection bind together: hard tasks get strong models, grunt work gets fast models — configured once, applied on every spawn.

Subagents' other pain point: context bloat, now handled officially

Another easily overlooked but equally important update in the same release: subagents now support automatic compaction (PR #64135). Midway through long sessions, subagents compress their own conversation history instead of degrading as context bloats, as before.

Anyone who has run multi-agent work knows this pain: you dispatch a subagent to "refactor the whole auth module," it grinds for 40 minutes — the first 30 are solid, the last 10 start reinventing wheels and dropping your earliest instructions, because those instructions got squeezed out of the context window. The old fix was manual: periodically ask for progress summaries and fresh sessions. Now the editor owns it at the infrastructure level, and long-task reliability moves from "good luck" to "guaranteed".

Taken together, Zed 1.22's theme is unmistakable: subagents are evolving from "one-shot prompt tools" into "orchestratable, long-running work units" — assignable models, self-managed context. That's the signature of the agent-infrastructure era: competition is no longer "whose model is stronger" but "who makes agent runtime mechanics more reliable."

Three things on the model side: BYOK, Grok 4.7, dynamic OpenCode fetching

The other half of the release is model onboarding, and it's equally information-dense.

First, BYOK support for GPT-6.1 Sol (PR #64968, community contributor ktKongTong). Bring your own OpenAI API key and use GPT-6.1 Sol inside Zed. The practical meaning for vibe coders: BYOK lets you ride your own API discounts, quotas, and billing instead of the editor vendor's resale pricing. As models commoditize, "where you buy tokens" is becoming as important a decision as "which model you use."

Second, Grok 4.7 joins the xAI and SuperGrok providers as the recommended default (PR #64567). xAI's model cadence keeps accelerating, and the recommended-default status signals the Zed team's assessment of its coding chops is in. Worth noting the move itself — the editor's default model pick is becoming an invisible battleground for model vendors.

Third, OpenCode's model list is now fetched dynamically at runtime (PR #64585) instead of relying on a bundled snapshot, with deprecated models auto-hidden (PR #64666). This fixes a real annoyance: vendors ship and retire models weekly while the editor's static list lags half a beat behind. After this change, "models visible in the editor" and "models vendors actually offer" are finally in sync — no more staring at a retired model name.

Workflow signals hiding in the "small" updates

Beyond the AI headlines, a few more 1.22 changes deserve vibe coders' attention:

Project settings support "..." placeholder inheritance merging (a set of PRs from community contributor porada). Monorepo subprojects can declare "inherit parent config and append" instead of copy-pasting it. Vibe projects that grow up tend to become monorepos (frontend + backend + shared packages); this kills config drift at the root.

The diff view opened via zed --diff gains a unified/split toggle (PR #64529). When the agent finishes changing code, you can flip between unified and side-by-side views while reviewing — a small review-UX thing, but review is exactly the least replaceable human link in the vibe workflow.

Terminal memory leak fixed: roughly 2MB retained per finished command (PR #64405, spotted by community contributor EcutAtom336). The terminal is the agent's hands and feet; agents run hundreds of commands a day, and leaks compound. The lesson: resource leaks in tools agents use heavily get amplified N-fold — "agent-friendliness" is becoming a real evaluation axis when picking tools.

One breaking change to note: the dock-close shortcut moves from cmd-w / ctrl-w to a unified ctrl-alt-w across platforms (PR #64515), because the old binding closed the dock instead of the tab. If shortcuts feel different after upgrading, that's intentional, not a bug.

Our take: the editor is becoming the agent's "operating system"

Plotting Zed 1.22 on the H2 2026 map, one trend keeps sharpening: the editor is morphing from "where you write code" into "the operating system agents run on" — spawn APIs, model scheduling, context management, terminals, diff review, every link of the agent workflow absorbed into the editor itself.

Two direct implications for vibe coders. First, choosing an editor is choosing agent infrastructure: model count no longer matters much; what matters is whether it makes your multi-agent division of labor stable, finely configurable, and cheaper. Next time you compare Cursor, Zed, and VS Code + plugins, put "subagent orchestration capability" on the scorecard — and weight it generously.

Second, start writing your AGENTS.md with an orchestration mindset: which tasks deserve strong models, which are fine on fast ones, where the checkpoints go on long tasks. Things that used to be improvised can now be configuration. Zed 1.22 hands you the knobs; the tuning is your craft.

Zed ships weekly, and the full 1.22 changelog is on the official releases page. If you're a Zed user, the first thing to try after upgrading is manually assigning a model to your next subagent and feeling the "selection bound to division of labor" workflow — you may never go back.