# how your agent thinks · InstaClaw docs

> your agent is a harness around several models, not one model: for every message it picks the model the message deserves, spends a share of your day's thinking on it, tells you if it ever has to swap, and lets you pick the model yourself whenever you want to.

people ask which model their agent is. the honest answer is several, chosen per message, and this page says how that choice is made so it never feels like a black box.

## which model answers?

your agent runs three tiers of model, and the tier is picked for each message by what the message asks for:

- **quick.** short questions, small edits, a reminder, a check on something. fast and cheap, and most messages are this.
- **standard.** the everyday workhorse: research, drafting, a plan, a file, most tool use. this is where your agent lives when it is working.
- **deep.** the flagship, for the messages that say so: be thorough, think it through, redesign this, build the whole thing from scratch, coordinate several agents. it is the most capable and the most expensive, so the picker reserves it for messages whose shape asks for it, and each plan has a daily number of these.

the picker reads the message itself: its length, the words that signal depth, whether you switched on deep research or web search, whether it is a background check rather than something you typed, and how much of the day's deep and standard budget is left. it runs in under a millisecond and it never sends your message anywhere to decide. if the picker ever fails, your agent falls back to its machine's default model rather than stalling.

## can i pick the model myself?

yes. the dashboard's model picker lets you set a model outright, and "automatic" hands the choice back to the picker. one model, the newest and most expensive, is never chosen automatically: it is used only when you pick it deliberately, because its cost should be your decision and never a side effect of how a sentence was phrased.

## what is a unit?

a unit is a share of your agent's daily thinking. each plan comes with a number of them, and a deep answer costs more units than a quick one, which is why the same plan can carry many small messages or fewer large ones. when the day's units are used, your agent pauses until they reset at midnight in your time zone and tells you so. [plans and billing](/docs/plans-and-billing) has the numbers; [heartbeat](/docs/heartbeat) explains why background checks draw from their own separate pool and never from yours.

## what happens when a model is unavailable?

models have bad hours. when the model your agent usually uses cannot answer, your agent does not go quiet: the reply comes from another funded model, and it says so at the top of the message, naming both, inside the platform notice that is never written by your agent. you are told about every swap, and a swap you were not told about is the one thing this design rules out. [when something breaks](/docs/when-something-breaks) has the rest.

## can i bring my own model access?

two ways. a key of your own from the model provider, which lowers the plan price and moves the daily allowance onto your key; or a ChatGPT subscription you already pay for, which can carry your agent's conversations instead of a key. [bring your own key](/docs/bring-your-own-key) has both, including exactly what each does and does not cover.

## where does the thinking happen?

on your agent's own machine, with the model called over the network for each message and nothing about you kept by the model between messages beyond what your agent chooses to write down. [your agent's computer](/docs/your-agents-computer) is the machine; [memory](/docs/memory) is what it keeps.

## what is switched on today?

the three tiers, the picker, the manual override and the daily units are on for every agent. the disclosed swap when a model is unavailable is built and not yet switched on; until it is, an unavailable model means a slower or missing answer rather than a substitute, and this page changes the day it is on.

## what next

- [plans and billing](/docs/plans-and-billing): the units per plan and what they buy
- [bring your own key](/docs/bring-your-own-key): your own key or your ChatGPT subscription
- [when something breaks](/docs/when-something-breaks): what a swap looks like when it happens