Adaptive Mode

  • Desktop app
  • All plans

What it does

Smart model routing from the composer. On a brand-new conversation, once you have typed a substantial draft and paused, AgentsRoom reads the start of the draft and suggests the cheapest model in your provider's line-up that can still do the task well: a light model for a typo fix, a rename, a translation or a summary; the balanced model for a pull-request review, a new endpoint or a refactor; the flagship for architecture, a security audit or a large legacy refactor. It tells you when your current model is already optimal. One click applies the suggestion: a running session is switched live and the choice is saved on the agent for next launch. The suggestion runs once per conversation by default, and the feature is opt-in. Two options sit on top of it: Every message (analyse each later message too, so the model can change from one message to the next, like a per-turn router) and Automatic (the agent switches to the chosen model before your message is sent, instead of a chip you click).

Video (1 min 26): Adaptive Mode, shown in the real app. Adaptive Model is turned on with Every message and Automatic, a new Claude Code agent on Opus switches to Haiku to fix a README typo, then goes back to Opus for an architecture question, without restarting.

Where to find it

  • The Adaptive Model chip above the composer, shown only once the feature is on and a suggestion exists ("Analyzing task…", "Switch to {model}", "Current model is optimal").
  • Settings > Optimization > Adaptive Model (in the AI agents group; it was in Composer until 2026-10-28, Composer now shows a link to it), On / Off (on by default on a new install since 2026-10-28, with a one-time notice above the composer, Keep it on / Turn off; installs made before keep their setting, off unless switched on; a Tutorial link at the right of the Adaptive Model title opens the video tutorial on YouTube), with its two options right under it: Every message (On / Off) and When a better-suited model is found (Suggest / Automatic). The settings menu also lists Adaptive Model and Every message as entries under Composer, which scroll straight to the section.
  • Per project: project settings > Agents > Adaptive Model, each choice on Default (follows the account) or set for this project only.
  • The Usage window, Optimize column, section Optimize your usage: an Adaptive Model card and an Every message card, both marked Recommended, each with an On / Off pill and a Turn on / Turn off link that opens the setting.
  • Scheduled agents and backlog tickets (Triggers window, an agent trigger's editor, a ticket's editor, and the phone's trigger and ticket editors): the Model card reads the task's own prompt and shows a short verdict under the model name ("Haiku is enough: fewer tokens", "Picked by Adaptive Model", "Adaptive Model awaits the prompt"). While a model is suggested, a Use {model} button sits right on the card (top right on the desktop, under the advice on the phone), so you accept without opening anything. Click the card itself: the top of its panel says "For this task, {model} is enough and saves tokens on your subscription." with a Use {model} button, or "Adaptive Model picked {model} for this task (it was on {previous})" in Automatic mode. With Adaptive Model off, a Pick the model from this prompt button does the same analysis on demand.
  • In Automatic mode, the chip above the composer shows Choosing the model… while the message waits, then Switched to {model} with thumbs to rate the choice. Hovering the chip shows "Adaptive Model picked {model} for this message (the agent was on {previous model})" and, once the API reported it, "Confirmed by the API: answered by …". The verdict usually arrives in under a second, so Choosing the model… is often barely visible.

How to use it

  1. Settings > Optimization > Adaptive Model > On. The setting says what is sent: your draft's first 500 characters go to AgentsRoom's servers.
  2. Open a fresh conversation with an agent and type your prompt. After a pause the chip analyses, then shows Switch to {model} or Current model is optimal.
  3. Click the chip to apply. If the CLI is running, the model is switched live with the CLI's own model command; in every case the agent's model is updated for the next launch.
  4. Rewrote the prompt? Click the refresh icon (Recompute the optimal model, costs one credit). Not interested? Dismiss suggestion.
  5. Want it on every message? Turn Every message on: each later message of the conversation is analysed the same way (from 20 characters).
  6. Want no click at all? Set When a better-suited model is found to Automatic: the agent moves to the model before your message is written. Rate the choice with the thumbs; they feed the routing quality.

Scheduled agents

The model of a scheduled agent is chosen once, in its editor, never at run time. With Adaptive Model on, the prompt is analysed after you change it (a 1.5 s pause, from 20 characters; opening an existing trigger costs nothing). Suggest shows the banner with Use {model}; Automatic sets the trigger's model itself, with any CLI, since only the saved model changes. A model you pick by hand afterwards stays until you edit the prompt again. With Adaptive Model off, nothing is sent until you click Pick the model from this prompt (one credit).

Settings

  • adaptiveModelEnabled (Settings > Optimization, "Adaptive Model"): on for a new install since 2026-10-28, off for installs made before unless switched on. While off, no chip and no network call.
  • adaptiveModelEveryMessage (Settings > Optimization, "Every message"): default off. Analyse every message sent, not only the first one of a conversation.
  • adaptiveModelApply (Settings > Optimization, "When a better-suited model is found"): suggest (default, a chip you accept) or auto (the model is switched before the message is sent).
  • The same three settings drive the advice on scheduled agents (Settings > Optimization shows the line "Scheduled agents follow this choice too").
  • adaptiveModelExcluded (Settings > Optimization > Adaptive Model, "Models it may choose"): per CLI, the models Adaptive Model never picks, e.g. { "claude": ["haiku"] } for Opus and Sonnet only. The AI classification of backlog tickets follows the same list. At least one model stays checked per CLI. Default: none excluded. Account only (synced, editable from the phone).
  • adaptiveModelContext (Settings > Optimization > Adaptive Model, "Context of the chosen model"): default (the chosen model keeps the agent's or project's context window) or max ("Always the largest": every model Adaptive Model or the classification applies gets its largest window, 1M on Claude models that offer it). Default default. Account only. In Automatic mode through the local server, the window stays the one the agent was launched with.
  • Backlog tickets block (Settings > Optimization, under Adaptive Model, since 2026-10-29; formerly Settings > Backlog): Automatic AI classification (backlogAutoClassify), the fields it fills (aiGenerateTitle, aiGenerateType, aiGenerateAgent, aiGenerateTags, aiGenerateModel, aiGenerateModelPick, aiGenerateEffort) and Always use these values (backlogClassifyPins). Same keys as before, Settings > Backlog keeps a link.
  • adaptiveTicketExplainSeen (internal, account): times the mascot bubble explaining a classified ticket was shown (3 = no longer shown). "Show the explanation on tickets again" under Backlog tickets resets it.
  • Scopes: account (synced across your computers and your phone) and project (project settings > Agents > Adaptive Model; "Default" = follow the account). The excluded models and the context option are account only.

Agent tools (MCP)

None of its own. settings_get / settings_set read or change adaptiveModelEnabled, adaptiveModelEveryMessage and adaptiveModelApply, at the global or the project scope, and adaptiveModelExcluded / adaptiveModelContext at the global scope.

Providers

Routes within Claude (Haiku, Sonnet, Opus), Codex (mini, baseline, flagship) and Antigravity (fast and pro models); on oh-my-pi it reads the models omp models lists for the user and maps them onto the same tiers, inside the account the agent runs on; for any other provider a generic rule applies. Every suggestion is validated against the models the provider actually offers, so it is always applicable. Automatic mode is available for Claude Code, Codex (local model router) and oh-my-pi (its /switch command, sent before the message); with any other CLI the chip suggests the model instead, even with Automatic selected. In Automatic mode, Claude Code and Codex agents are started connected to a small local server inside AgentsRoom: the model of each message changes there, with no relaunch, nothing typed in the CLI, and your account login untouched (the Claude account's default model is not changed either). This applies to agents started after Automatic is turned on; restart an agent that was already running. Agents on a remote SSH host, or whose CLI already goes through your own proxy or API endpoint (HTTPS_PROXY, ANTHROPIC_BASE_URL, Bedrock, Vertex, OPENAI_BASE_URL, a Codex model_provider), are never rerouted: they switch with the CLI's own model command (Claude, Copilot, Grok, Kimi, Devin, Pi, OMP) or, for Codex, by resuming the same conversation on the new model. In Suggest mode, applying a chip still uses the CLI's own model command.

Mobile

Activity tab > Optimize (the third segment, next to Active agents and Usage; since 2026-10-29 it has its own segment, so the Usage screen shows your CLI consumption straight away) > Optimize your usage: Adaptive Model On / Off, Every message, Suggest / Automatic, Models it may choose and Context of the chosen model (account level; project overrides are set on the desktop). In Suggest mode the chip shows above the phone's composer too, with the thumbs. In Automatic mode, messages sent from the phone are routed by the desktop before delivery.

Limits

  • Monthly suggestions: 50 on Free and Plus (500 until 2026-10-28), 10,000 on Pro; 10 without an account. One credit per analysed message (once per conversation by default, every message with the option); each manual recompute costs one credit. When the allowance is spent, messages go out on the current model.
  • Only the first 500 characters of the draft leave the machine, and only when the feature is on.
  • In Suggest mode it never switches the model without your click. In Automatic mode it switches, and a message waits at most 4 seconds for the verdict; it never blocks sending the prompt.
  • Without the Every message option, auto-analysis triggers only on a conversation with no message sent yet.
  • A message sent while the agent is still working on a turn is not routed (the switch could only apply after the turn).
  • Switching model mid-conversation drops the prompt cache of the previous model: the first message on the new model re-reads the history at full price.

Common questions

  • How do I turn it on or off? Settings > Optimization > Adaptive Model. On by default on a new install; while off nothing analyses your draft.
  • Does it switch automatically? Only if you set When a better-suited model is found to Automatic. By default it suggests and you apply with one click.
  • Why no chip on my second message? By default the analysis runs once per new conversation. Turn on Every message to analyse each message, or use the recompute button.
  • Can AgentsRoom change the model at each message, like the Jev router? Yes: Adaptive Model > Every message > On, and Automatic if you want no click. It runs inside AgentsRoom, with nothing to install: no plugin, no extra CLI, no API key.
  • How much does it save me? The Usage panel (desktop and phone) shows under Adaptive Model how many of this month's messages it moved to a cheaper model.
  • How do I know the model really changed? In Automatic mode the chip above the composer says Switched to {model}, the agent's model pill in its header follows with a short animation (also when you change the model by hand), and a check badge appears once the API itself reported which model answered (hover: "Confirmed by the API: answered by …"). Codex's own footer keeps showing the model it was started on.
  • Does it work behind a corporate proxy? Yes: the switch goes through the same network stack as the rest of AgentsRoom (system proxy, PAC, certificates installed in the system, and the proxy set for the provider in Settings > AI providers). If you set a proxy or a certificate file yourself in an agent's environment variables, AgentsRoom does not reroute that agent and it switches with the CLI's own model command instead.
  • How do I make my agent use Haiku for small fixes and Opus for big questions? Turn Adaptive Model On in Settings > Optimization, then Every message On and When a better-suited model is found on Automatic, and start a new Claude Code or Codex agent. A typo fix goes to the light model, an architecture question to the flagship, with no restart. Video: https://youtu.be/wFQySlynVug
  • Is there a video that shows it? Yes, about a minute and a half, filmed in the real app: https://youtu.be/wFQySlynVug
  • Does Codex restart to change model? No, in Automatic mode a Codex agent started after the option was turned on changes model without restarting.
  • Can I turn it on for one project only? Yes: project settings > Agents > Adaptive Model, set On there and leave the account Off.
  • Can it pick the model of my scheduled agents? Yes. Open a trigger's editor: in the Model card, Adaptive Model says which model is enough for the prompt and how it saves tokens on your subscription. In Automatic mode it sets the model itself each time you change the prompt; otherwise click Use {model}. With Adaptive Model off, click Pick the model from this prompt.
  • Is it on by default? On a new install, yes (since 2026-10-28): a notice above the composer says so and what is sent, with Keep it on and Turn off. Installs made before keep their own setting (off unless switched on), and a new computer never switches it on for an account that already exists.
  • Does it work on backlog tickets? Yes, the same way: open a ticket's editor (Execution section), the Model card reads the ticket's title and description. Click the card for Use {model}, or Pick the model from this ticket when Adaptive Model is off.
  • Can I stop Adaptive Mode from ever picking Haiku? Yes: Settings > Optimization > Adaptive Model > Models it may choose, uncheck Haiku under Claude. Adaptive Model and the backlog classification then only pick among the checked models.
  • Can the chosen model always run with the 1M context? Yes: Context of the chosen model > Always the largest. Every model Adaptive Model applies (composer chip, Automatic, scheduled agents, tickets, classification) gets its largest window.
  • Why did my new ticket's agent, model or effort change by itself? The automatic classification (part of Adaptive Mode) filled them from the description. The first three times, a bubble with the mascot says so above the ticket fields, with Got it, Configure (opens the Backlog tickets block of Settings > Optimization) and Turn off (desktop). Each field can be switched off there.
  • The backlog AI classification also picks a model: which one wins? A model imposed for tickets in Settings > Optimization > Backlog tickets > Always use these values (account or project) always wins: that is how to keep Opus on tickets and Adaptive Model elsewhere. Otherwise, with Adaptive Model on, Adaptive Model chooses (Suggest or Automatic) and the classification only fills title, type, agent, tags and effort. With Adaptive Model off, the classification picks the model as before.
  • Is my prompt private? Only the first 500 characters are sent to compute the suggestion, and only when the feature is on. Nothing is stored with your code.
  • Difference with AI Suggestion? AI Suggestion decides the model before the session opens (intent screen); Adaptive Mode suggests from the draft after the CLI has started. Both share the same sizing policy.