How to have an AI agent plan and run your generations

Agent mode is a chat where every turn drives the platform's own tool registry. You describe an outcome; the model restates what it heard, proposes a plan of concrete tool calls, prices each step in credits against the wallet that would pay for it, and then waits. Nothing runs until you press Run. This page covers what the thirty-five tools reach, how the quote works, and where the Beta stops.

What agent mode is

Agent mode used to be a per-message toggle buried in the attachment menu: it reset on reload, was invisible from another device, and a conversation full of plans looked exactly like a conversation full of prose. It is a section now. Every conversation started in Agent runs with the generation tools on, permanently, and those conversations are filed separately from ordinary chats.

What the model is handed is the platform's own tool registry — the same tool definitions the platform exposes over MCP. Inside a turn it can ask which models this account may run, price a run before proposing it, generate images, video and audio, pull a video in from a link, describe and transcribe what you attached, cut moments out of a long video, write copy, save results into your library and notes into a brain, and stage and start a LoRA training run.

Which plan: The Agent section is Creator, Pro and Ultra, badged Beta. Below Creator the section is still listed and still explains itself rather than disappearing — it just does not run. These tools spend real GPU time on a single workstation, which is the whole reason the gate exists.

It quotes before it spends

This is the part worth reading twice. Every tool in the registry that can spend credits automatically gains a dry_run parameter — the tools do not each implement one, the registry adds it. Called with dry_run: true, the registry answers with the price and never reaches the tool's handler at all.

What a dry run answers with, before anything has run
FieldWhat it means
creditsThe number that will be charged if this step runs
walletWhich wallet pays — green for models on our own GPUs, gold for hosted providers
quotableFalse means the step cannot be priced and must not be run as it stands
priced_forThe basis of the number in words, such as the clip length and resolution a video price was worked out from
reasonWhy it could not be priced, when quotable is false

The prices in that answer are not a parallel estimate. The quoting module calls the same cost functions the generate routes bill with and adds no prices of its own, which is the only way a quote and a charge can be guaranteed to be the same number.

A step the engine cannot price comes back marked unquotable with the reason attached, and that reason is shown on the step's own tile. An unpriceable step stops the run rather than riding an invented number — a missing input, a model your plan cannot run, or a training job whose hardware budget could not be read.

The plan is a rail you can edit

A proposal renders above the composer as tiles grouped by stage, not as a wall of JSON. The stages are the shape of a run — analyse, prepare, create, then whatever the created thing turns into: derive, animate, sound, text, train, library. Work derived from an earlier step nests under its source, so a ten-step plan reads as a tree instead of a list.

Steps chain by position. When a later step should run on the output of an earlier one, the earlier step is referenced as a {{step N}} token inside the later step's arguments — the enhanced prompt into the image call, the image into the video call. Numbering is per proposal and can only point backwards.

The rail is editable before it runs and after. Switch a step off and it never runs. Every Run press is recorded against the plan, so a plan can be run in passes: generate the images, look at them, run the video step, then re-roll only the one image you did not like. The newest result for a step wins and the older ones stay as that step's versions.

Asking for a change revises the plan rather than replacing it. Steps already run are frozen and re-listed first so their numbers hold and their outputs stay referenceable; only the not-run steps move.

How to run your first plan

  1. Open the Agent section. Go to /llms and pick Agent in the rail. A conversation started there keeps its tools on for good — you cannot half-leave agent mode by reloading the page.
  2. Describe the outcome, not the tool. "Three product shots of this bottle on a pale background, then an 8-second clip from the best one" is a better brief than a list of model names. It picks the models and you overrule it where you disagree.
  3. Read the read-back first. Before proposing anything the agent restates which of your words became which action. That read-back runs immediately, spends nothing, and is the record the plan is then checked against — if it misheard you, say so before a plan exists.
  4. Check every price. Each tile carries its own credit number and the wallet that would pay it. A tile with no number is a step that could not be priced, and it names the reason on its face.
  5. Trim the plan. Switch off anything you did not want. Approving four steps out of eight is a normal way to use this, not an edge case.
  6. Press Run. Only now does anything spend. Results land on the plan's own tiles and in your assets library.

Open the Agent section — Opens the chat page. Agent is the second section in the rail; nothing runs until you press Run on a plan.

What the thirty-five tools reach

None of the tools is owner-only, so a Creator account sees the whole registry. They fall into eight groups.

The tool registry, grouped, with what each group spends
GroupToolsWhat it spends
Look before it leapslist_models, get_model_schema, describe_capabilities, get_wallet, quote_run, read_requestNothing
Make pictures, video and soundgenerate_image, generate_video, generate_audio, tts, sfx, list_voicesYour wallet, quoted per step first
Read what you gave itanalyse_inputs, describe_image, describe_video, probe_video, detect_scenes, transcribeSome steps spend; each is quoted
Bring video in and cut it upingest_video, import_video_url, extract_momentsGold, and metered against your plan's analysis minutes
Plan and writeenhance_prompt, plan_movie_scenes, plan_deliverables, write_copyPrompt enhance costs 0.2 credits; the planners are free
Your library and your brainslist_assets, get_asset, save_asset, brain_recall, brain_noteNothing; brains need a vault, which Creator has
Train a LoRAlist_lora_options, create_lora_dataset, train_lora, get_training_status1,000 credits per GPU-hour, charged up front at the estimate
Watch a jobget_job_statusNothing

What a run actually spends

The agent spends your wallet at exactly the prices the rest of the site charges, because it calls the same pricing code. Generations on our own GPUs start from ⚡1 and are paid in green; hosted models from Google, ByteDance, OpenAI and the rest are paid in gold, which is bought rather than refilled. The wallet column on each tile tells you which of the two a step will draw on before you approve it.

What it costs: Own-GPU prices start at ⚡1 and are a floor, not a fixed number. They are set at runtime, can carry a demand multiplier when the render queue is deep, and when the local lane is unavailable the same model is served over a paid API and billed in gold instead. The quote on the tile is priced at the moment you see it.

LoRA training is the one step that charges before it produces anything. train_lora prices the run from the model, the step count and the size of the staged set at 1,000 credits per GPU-hour, charges that estimate up front from the gold wallet, and refunds it in full if the run never starts. It is also the only place LoRA training is reachable outside admin hands — see how to train a LoRA.

A daily chat allowance running out mid-plan does not end the run either. The refusal buys the smallest useful allowance from your wallet, green credits first, and the turn continues. If neither wallet can fund it, the refusal stands and says so plainly instead of failing silently.

Where the Beta stops

Watch out: A plan is only as good as what it was given. Attach the reference images and the source video before asking for the plan — a step that has to invent its input is the step most likely to come back unquotable.

Will it spend my credits without asking?

No. A plan is a proposal: every step is priced, the wallet that pays is named on the tile, and nothing runs until you press Run. Underneath that, every spending tool has a dry-run mode that returns the price without touching the tool's handler at all.

What plan do I need?

Creator, Pro or Ultra. The section stays visible on Free and Starter and explains what it does, but it will not run there. It is badged Beta on the plans that can run it.

Can the agent really train a LoRA?

Yes, and it is the only way to reach training outside admin hands. It stages a dataset, prices the run at 1,000 credits per GPU-hour, charges that estimate up front from the gold wallet, and refunds it in full if the run never starts. Ask for a dry run first and read the number.

What happens if a step cannot be priced?

It comes back marked unquotable with a reason on its tile, and it must not run as it stands. Fix what the reason names — usually a model your plan cannot run, or a missing input — and ask again.

Can I run only part of a plan?

Yes. Switch steps off before pressing Run, or run the plan in passes: generate the images, look at them, then run the video step. Re-rolling one step leaves the rest alone, and the older result stays as a version of that step.

Does it work with my own agent or desktop client?

No. Agent mode runs in the Agent section of the chat page on OpenModels, against your signed-in account.