Shared context for AI coding agents beats better tooling

The instinct when adopting AI coding agents is to optimize the agent. Compare models, tune prompts, argue about which editor has the better completion, and treat the agent as the thing that determines how fast the team moves. Then the commits go up and the product does not.

The team building Upsun Dispatch™ took a different route, and the result is worth copying. They did not find a better agent. They gave every person and every agent on the team the same context to work from: one repository, one shared set of documents, one glossary everybody holds to. Speed followed from that.

The agent was never the bottleneck

Patrick Dawkins, principal engineer at Upsun, describes the constraint plainly in the latest Product Highlights episode: the team decided to build something ambitious in a short space of time, and then had to work out how to actually sustain that pace.

The tooling part was the easy part. "We are highly AI enabled in the team," Patrick says. "A lot of that is happening naturally, so everybody's using their coding agents. We have our review agent running." Nobody had to be talked into it.

But individual adoption is where the gains tend to stall. Upsun has written about this before: the constraint was never typing. An engineer who gets fast on their own machine, with their own framework and their own accumulated context, produces work the rest of the team still has to understand, review, and integrate. Multiply that by several agents running in parallel and the review queue absorbs every hour the agents saved.

One repository, one set of documents, one vocabulary

What changed at Upsun was the substrate the agents run on, not the agents.

Everyone works in the same monorepo. Documentation is shared across the whole team, and not just engineering documentation. "I think this has come from marketing and product and engineering, all together," Patrick says. The docs cover planning the work and planning the long-term vision, so the reasoning behind a decision sits next to the code that implements it.

That has a second effect that is easy to miss. Every document written for a human is also a document an agent can read. As Patrick puts it: "As well as the team all having access to the same things and the same vocabulary, all the agents also have access to those docs."

The access is deliberately wide. The team and the agents can both reach the back end, the front end, the infrastructure code, and the Terraform configs that build the infrastructure. An agent asked to change a service is not guessing at how that service is deployed.

What happens when non-engineering context stops hiding

The clearest illustration came from outside engineering. Greg Qualls, senior director of product marketing and content at Upsun, has spent almost five years at the company and describes product marketing work as something that historically happened out of sight of the codebase.

Putting it in the shared repo changed who could use it. It was no longer his context, or even his agents' context. As Greg puts it: "It's available to you and your AI agents."

The practical difference shows up in the small questions. The kind of question that used to wait for a meeting, or wait for the right person to be online, now has an answer sitting in the repo that either a human or an agent can find and act on. Multiply small unblocked decisions across a team working at pace and it stops being a small thing.

This is the same conviction that shapes Upsun Dispatch as a product: the whole team ships, not just engineers. Product, design, and security take part in a workflow directly, following runs and approving gates without touching code.

A glossary is infrastructure now

The most transferable idea in the conversation is also the least glamorous. Patrick is strict about naming, and for a specific reason.

When you build something new, you have to name things, and the candidate names are frustratingly close together. Activity. Task. Job. Execution. Worse, they cannot always be collapsed into one term, because a tool in the stack may already claim one of them. Temporal, for example, brings its own vocabulary, and the team has to keep meaning aligned across every layer above and below it.

Get that wrong and the cost compounds twice over. "Otherwise there's going to be confusion for people and confusion for our agents, forever," Patrick says. "So I am quite pedantic with our glossary."

Vocabulary drift has always been expensive. One person using a term slightly wrong in a design conversation can send a week of work in a direction nobody intended, and the mistake usually surfaces late. Agents make that worse in a particular way: they are fluent, confident, and entirely dependent on the words in front of them. An agent handed two definitions of "task" usually doesn't stop to ask which one you meant. It picks one and writes code.

So the glossary stops being documentation hygiene and becomes part of how the system behaves. It is worth the pedantry.

Then get the pipeline out of the way

Shared context raises how much work the team can start. It also raises how much work arrives at review, which is where the pressure lands next.

Patrick's answer is unglamorous and correct: "What I've been trying to work on is making all our CI very efficient, just to get out of the way, and help us constantly review and merge PRs." Not faster review by cutting corners. Faster review because the pipeline stops being a reason to wait.

That ordering matters for anyone trying this. Shared context without a fast pipeline just relocates the queue.

Where Upsun Dispatch fits

None of the above requires a product. A team can put its docs in the repo, write a glossary, and tighten its CI starting Monday. What is hard to do by hand is everything after that: running agents somewhere other than a laptop, making a workflow run the same way twice, keeping a record of who approved what, and knowing what a run cost.

That is the gap Upsun Dispatch is built to close. The founding idea is that the workflow is the primitive, not the agent. Agents run in isolated sandboxes rather than on personal machines, each workflow is a defined sequence of agent and human steps, human gates require a real decision instead of a notification, and every run is logged with its cost attached. Paired with Upsun Cloud preview environments, each change also gets a byte-for-byte copy of production to be tested against.

The reason the internal story is relevant is that the product encodes the same belief. Shared context, an agreed vocabulary, and a process the whole team can see are what turn capable models into a team that ships. The agent is the easy part.

Upsun Dispatch is currently open to a founding design partner cohort, aimed at engineering teams of 50 to 500, with a full public launch planned for September 2026. Design partners work directly with the engineering team and influence which workflows ship first. 

You can join the waitlist or watch the full conversation with Patrick at here.