Before your AI bottleneck gets worse: what to put in place now

Your engineers have agents running. Not one agent, but several, spread across the team. Some run in a terminal on a laptop, some are wired into your CI jobs, and some live inside whatever coding tool each person prefers. Each one got set up separately, by whoever needed it, in whatever way worked that week.

That is the state most teams are in right now. Code stopped being the slow part a while ago. What is slow now is everything wrapped around it, and the thing almost no one has stopped to decide is where all of this actually runs. It has been accumulating by default, one agent and one CI job at a time. No one chose the arrangement. It simply grew.

The layer you did not know you were choosing

Every team running agents is standing on a platform layer: the thing that decides where agents execute, which models they use, how they reach your infrastructure, and who can see what they did. You are using one, whether or not you picked it.

There are two common ways to end up with one without choosing it.

The first is to extend what you already have. Agents get bolted onto your existing CI, the same system that runs your builds and tests. It works at first. But the CI vendor built that system to run pipelines, not to govern a team of agents, and extending it means inheriting its assumptions: the models it makes convenient, the compute it runs on, the deployment path it expects, and its idea of what an agent workflow should look like. You did not evaluate those assumptions. They came with the tool you already had.

The second is to do nothing deliberate at all. The layer assembles itself from whatever each engineer wired up locally: one person's terminal setup, another's script, a third's CI experiment. There is no layer so much as a pile of them, and no one can see across it.

Both are choices. Neither was made on purpose. That is the problem. The platform layer gets decided either way, and inheriting one is not the same as choosing one.

What the layer decides

A platform layer is the set of decisions that sit under every agent your team runs: which model answers a given task, what compute the agent runs on, how it reaches your code and infrastructure, and what record it leaves behind. 

Whether you chose that layer or inherited it, it quietly settles three things you will care about later. It is worth seeing them concretely.

1. Which models can you use

Frontier Labs is burning cash, and its pricing will move. Tie your layer to one provider, and that change is yours to absorb. A layer that lets you point each task at the model that fits and swap providers when the economics change, without rebuilding what sits on top. Most coding tasks do not need the largest model anyway, and a smaller one, configured well, costs far less.

2. Where it runs

An inherited layer runs where the vendor runs, on their terms. A deliberate layer runs where your infrastructure already is. Upsun Dispatch™ is a standalone product for exactly this reason: it does not require Upsun Cloud to run your agents and your development lifecycle. You bring your own arrangement rather than adopting someone else's.

3. What you can see and control

This is one that hurts most when it is missing. Agents scattered across laptops and CI jobs leave no one able to see across the organization: not why one agent spent thousands of dollars, not what a change touched, not who approved it.

 A deliberate layer makes agent work observable and governable, with human gates, cost attributed to the workflow that spent it, and an audit trail of prompts, diffs, and approvals. Work you cannot account for is a liability that grows with every agent you add.

Where the inherited layer is enough

Not every team needs to make this decision yet, and it is worth being honest about that.

If you are building a throwaway project, a one-off internal tool, a fun web page, or anything you will not maintain, the code barely matters, and neither does the layer under it. Point an agent at it and move on. The same is true for a team still early in AI adoption, mostly using autocomplete, not yet running agents in parallel, not yet feeling the sprawl. There is nothing to govern, so there is nothing this decision saves you from.

What deliberate looks like

Choosing the layer on purpose means running your agents on something built to govern them, rather than something built for another job and stretched to fit.

That is what Upsun Dispatch is: a platform for governed collaboration between AI agents and cross-functional human teams. It sits across from your agents and models rather than being another agent in the pile. It runs standalone, where your infrastructure already is, and lets you choose which model handles which task. And it brings the visibility and control that a stretched CI system was never designed to give you, built on structured workflows, isolated sandboxes, human gates, and a full audit trail.

The point is that the layer was chosen for the job by you, instead of being inherited from a tool that happened to already be there.

What to put in place now

The layer gets decided either way. The only question is whether you decide it or inherit it. Deciding it now is cheap and reversible. You put a chosen platform layer under your agent work while the sprawl is still small enough to redirect. Inheriting one and unwinding it later is neither cheap nor reversible: you are re-platforming agent workflows your whole team now depends on, under the models, compute, and deployment assumptions of a vendor who made those choices for you.

The straightforward way to make the deliberate choice real is to run one workflow on a layer you actually picked.  

Start a workspace in Upsun Dispatch and run a workflow through it. You will see, in your own work, what it means to own the layer instead of inheriting it.

Frequently asked questions (FAQs)

What is a platform layer for AI agents? 
It is the set of decisions sitting under every agent your team runs: which model handles a task, what compute the agent runs on, how it reaches your code and infrastructure, and what record it leaves behind. Every team running agents has one, whether they chose it deliberately or inherited it by extending their CI or letting setups accumulate across laptops.

Can I just run AI agents on my existing CI pipeline? 
You can, and many teams do, but CI systems were built to run pipelines, not to govern a team of agents. Extending CI means inheriting its assumptions about models, compute, and deployment, and it was never designed to give you cost attribution, human approval gates, or an audit trail across agents. It works early on and gets harder to unwind as agent use grows.

Why does it matter which model my agents use? 
Frontier model pricing changes often, and tying your setup to one provider means absorbing those changes directly. A layer that lets you match each task to a suitable model and switch providers when the economics shift avoids a costly rebuild later. Most coding tasks do not need the largest, most expensive model.

Do I need Upsun Cloud to use Upsun Dispatch? 
No. Upsun Dispatch is a standalone product. It runs your agents and development lifecycle without requiring Upsun Cloud, so it works where your infrastructure already is rather than forcing you onto a specific host.

When is an inherited platform layer good enough? 
For throwaway projects, one-off internal tools, or anything you will not maintain, the layer underneath barely matters. The same applies to teams early in AI adoption who are mostly using autocomplete and not yet running agents in parallel. If there is nothing to govern, this decision does not save you from anything yet.