There is a stubborn belief in agentic development: "when in doubt, give the agent everything". The whole repository, the whole history, the whole documentation. The intuition feels safe - and it is an expensive mistake. Bloated context means a token bill that grows on every single run, and often a worse answer, because the agent drowns in noise.
At Survol we start from a simple, almost accounting principle: tokens are the customer's money. We do not resell tokens; the subscription stays yours. So our job is not to make you burn more of them, but to make you burn as few as possible for the same quality. Here is how.
The right context is not the biggest one. It is the smallest one that is enough.
1. Compose the context, don't dump it
When you approve a feature, Survol does not throw the whole repository at the agent. It assembles a targeted context from three ingredients, and three only:
- the approved spec of the feature - what you want, written properly once;
- the decisions already settled that relate to it - so nobody reopens old debates;
- the files actually concerned - not all the code, the ones that matter.
That trio changes everything. Where a naive prompt sends 80,000 tokens of context "just to be safe", a composed context sends a fraction of that - and the agent aims better because it does not have to guess what is relevant.
# Context composed by Survol from the approved spec context: feature: "Promo codes & gift cards" version: "v2.0" decisions: ["cannot be combined with sale prices"] files: ["src/Pricing/Promo.cs", "src/Checkout/Cart.cs"] model: "selected automatically for the task"
2. The right model for the right task
Not everything deserves the most powerful model. Renaming a variable, adjusting a label, writing a simple unit test: a lightweight model does the job for a fraction of the price. A multi-file refactor or a nasty debugging session: that is when you bring out the heavy artillery. Survol picks the model automatically based on the nature of the task, instead of letting you pay premium rates for everything, all the time.
Order of magnitude: between a context composed as tightly as possible and a model matched to each task, we aim for roughly a third fewer tokens over a typical development cycle - with no perceptible loss of quality. (That is an internal target, not a contractual guarantee: it all depends on your code.)
3. Measure so you can decide
You only optimise what you measure. Survol tracks consumption per session, per version and per feature - not just per person. You see which feature cost what, and you finally connect spending to what was actually delivered. A feature burning tokens without moving forward stands out at a glance, and triggers the right question: badly cut context? ambiguous spec? poorly split task?
What it looks like day to day
- You approve a spec; Survol composes the matching minimal context.
- The model is chosen for the task, without you having to think about it.
- Consumption is displayed, attached to the feature and to the version.
- You arbitrate on numbers, not on an end-of-month hunch.
Optimising tokens is not about being stingy: it is about respect. Your subscription is a finite resource, and so is your time. A tailored context gives both their due - and, incidentally, makes the agent better. The smallest context that is enough almost always beats the biggest one.
