Product·Auto Model Mode

Pick Auto.
Continuum picks the team.

Auto Model Mode is a BYOK router across planner, executor, and verifier roles, on the hosts you already run. It triages the prompt first, picks each role from scored model families, stamps every stage with the model that actually ran it, asks before the executor writes, and treats Cancel as the brake. Honest v1: it is not quota-aware yet.

Download for Mac Read the docs
CodeUsageSettings
auto
São Paulo · AutoBYOK · your providers
planner fable-5 score 0.91
executor gpt-5.5 score 0.87
verifier grok score 0.84
role routingnot quota-aware · v1
01 · Roles

Planner. Executor. Verifier.

Auto doesn’t mean “mystery model.” It means Continuum assigns roles across the providers you connected: one model plans, one implements, one checks. Each hop can land on a different host in your fleet.

Auto is not a separate provider. It sits in the model picker next to your real models, and before you start anything it tells you which model would plan, so the first thing you learn about a run is not a surprise after the fact. It only appears when the host supports routing and at least one lane is genuinely usable: enabled, authenticated, and reachable.

A lane is a whole vendor rail, not a single model: Claude, Codex, Gemini, Cursor, Grok, OpenCode, OpenRouter, and Z.ai all enter the pool once you sign in. You still own every login. Continuum only routes among keys you already pay for, which is the difference between this and picking one agent and living with it.

three roles cross-host BYOK only
01 plan

scope · steps · risks

→
02 execute

edits · tests · tools

→
03 verify

review · gate · ship

02 · Triage

Not every prompt deserves a pipeline.

Before it routes anything, Auto reads the prompt and decides how much machinery the work is worth. A three-model plan-execute-verify run on “fix this typo” is a way to spend four times the money for a worse answer, so a small prompt gets one fast model and nothing else.

01

Score the request

Length pushes the score up past 400 and again past 1,200 characters. So do architecture words like refactor, migrate, protocol, and end-to-end, three or more file paths, a numbered shape, and a large repository. Scoped-fix words like typo and quick fix pull it back down, and a bare question drops it further.

02

Fast path or full pipeline

Above the threshold, Auto runs the full planner → executor → verifier pipeline. Below it, a single fast executor takes the turn: the quickest model that still clears an intelligence floor, never just the cheapest one available. If nothing clears the floor, Auto says so in the reason and falls back to the strongest model instead.

03

Notice taste-sensitive work

Prompts mentioning UI, UX, design, copy, layout, animation, naming, API design, or polish are flagged taste-sensitive, and the executor pool is filtered to models that score well on taste before intelligence gets to break the tie. When no model clears the taste bar, the reason string admits it rather than quietly downgrading.

Code
scores
fable-5
0.91
gpt-5.5
0.87
composer
0.72
grok
0.84
scores · task fit · not hidden magic
03 · Score-based routing

Picks you can inspect.

Selection is score-based, task fit across the models available in your catalog, not a black-box vendor ranking. Every model family carries five numbers: cost, intelligence, taste, speed, and agentic ability. The table ships with the app and is versioned, and the version that produced a decision is recorded on the run, so a routing choice you disagree with can be traced to the exact snapshot that made it.

Each role reads those numbers differently. The planner takes the highest intelligence, then taste, then agentic. The executor weights intelligence three times against agentic ability, because an executor that reasons well but can’t drive tools is useless. The verifier takes intelligence first and prefers a different provider from the executor entirely. A model comparison is the same exercise done by hand, once, instead of per prompt.

When one model family is reachable on more than one rail, Auto prefers the rail that costs you less: a first-party subscription you already pay for beats a metered API key, which beats an aggregator. Availability still matters: a provider that isn’t connected or is disabled never enters the pool. On the economics of that choice, see AI coding pricing compared.

04 · Role stamps

See which model ran.

Transcript steps carry role stamps: planner, executor, or verifier, plus the model id. A multi-model run never becomes an unattributable blur. Each stage records the model, the provider, the rail it billed, the effort level, when it started and ended, and what it cost. Cost and tokens still roll up honestly under each provider in analytics.

Stamps are written from what actually happened, not from the plan. Every role assignment carries up to three ranked fallbacks, and when one fires, the stage is re-stamped with the model that really ran and a record of what it fell back from. A run where Claude was busy and Codex picked up the executor role reads that way in the transcript, permanently.

If you run more than one subscription per vendor, routing picks the rail and a second pass binds your preferred account for that vendor, so an Auto run bills the subscription you pinned rather than whichever login answered first. When something goes wrong, you know which role to re-run, not “the AI.”

Code
stamps
planner · fable-5

4-step migration plan ready for approval.

executor · gpt-5.5

Applied parser isolation; tests running.

verifier · grok

Diff review: +212 −96 · no secrets in patch.

05 · Verify and fix

The verifier has to say the word.

A second model reading a diff and saying “looks good” is theatre. Auto’s verifier stage is held to a machine-checkable answer, and the failure direction is deliberately the safe one.

01

A verdict, not a vibe

The verifier is instructed to end its reply with VERDICT: PASS or VERDICT: FAIL on its own line, then list problems as numbered findings. The parser is lenient about markdown and casing and always takes the last verdict in the reply, so a model that reasons its way from FAIL to PASS resolves to its final call.

02

Unreadable means fail

If no verdict can be parsed, the run is treated as failed. Never as passed. A verifier that rambles, gets cut off, or ignores the format cannot accidentally wave work through, which is the only tolerable default when the alternative is unverified code with a green badge on it.

03

The fix loop is bounded

A FAIL sends the findings back to the executor as a fix round. That happens at most twice. After the second attempt the run completes with warnings and the outstanding findings stay on the record, because a router that grinds forever on a problem it can’t solve is worse than one that hands the problem back to you.

Auto planask first
Planner proposes steps
You approve before execute
cancel the brake · stops the run
06 · Ask-first & cancel

Auto never means unsupervised.

Ask-first plan approval keeps a human gate before the executor writes. Same plan card, same Approve path as manual plan mode in the Code workbench. A run announces exactly where it is at all times: triaging, routing, planning, awaiting approval, executing, verifying, fixing, summarizing, then completed, completed with warnings, failed, cancelled, or suspended.

Cancel is the brake. It interrupts the active role immediately. Auto is a router with a kill switch, not a runaway swarm. Details on plan & review.

Every transition writes a durable receipt on the host: the phase, the session id for each role, the fix iteration, the triage result, the routing decision, and the approved plan text. Restart the daemon mid-run and it rehydrates at the last stage boundary instead of starting over. A run whose receipt predates the fields needed to continue is suspended rather than guessed at, which is the honest outcome. This is what makes long agentic runs survivable.

07 · Honest limits

Not quota-aware in v1.

Auto does not yet read live rate-limit gauges when picking models. It won’t automatically dodge a 95% Claude window. Watch the menu-bar gauges yourself, pin accounts, or pick models manually when a window is hot. If you don’t know where your windows stand, start with how the usage limits actually work.

We’d rather say that up front than pretend the router is smarter about caps than it is. Quota-aware routing is a later chapter. Two other things Auto does not do: it never silently caps spend, and it never runs unattended for hours without a gate. See gauges and multi-account, or cut the cost per run first.

v1 does
  • role routing across BYOK models
  • score-based selection you can see
  • stamps · ask-first · cancel
v1 does not
  • quota / rate-limit avoidance
  • silent spend caps
  • unattended multi-hour fleets
08 · Questions

The honest fine print.

Short answers here, long answers in the docs.

No. Auto routes among providers you already connected (BYOK). Continuum is not selling a mystery model under the Auto name. If you want a single-vendor router instead, that is the model Cursor and Codex ship, and the comparisons spell out the trade.

Only as a last resort, and the run says so. Auto prefers a verifier on a different provider entirely. If nothing cross-provider is available it takes a different model on the same provider; if you have exactly one lane connected it falls back to the same model and records that the check was weak. Two lanes is the minimum for a real second opinion.

Each role carries up to three ranked fallbacks resolved at routing time. When one fires, the stage is re-stamped with the model that actually ran and the one it replaced, so the transcript shows the substitution rather than hiding it. If no lane is usable at all, the run fails instead of silently picking something you never connected.

Cancel stops the active work. You can continue the session with a manual model pick for the next turn. Auto doesn’t lock you out of the composer.

Yes when the host can spawn the selected providers. Role work still runs on your hardware; controllers only steer.

Not in v1. That limitation is intentional product honesty. Use gauges, multi-account pins, or manual selection when a window is hot.

09 · Begin

Let Auto draft the team.
You still approve the plan.

Connect two providers, start a Code session, pick Auto.

Download for Mac Read the docs

see also · code · providers · plan & review