The Hidden Tax on Your AI Agents

AI coding agents solved the generation problem. Code ships faster, more services get built, output goes up. But something else happened that nobody planned for.

Agents started moving faster than anyone could review. They produced work that was technically correct and operationally wrong — because they had no idea what they were working inside. They couldn't see who owns the upstream service, didn't know the dependency graph, and had no awareness of why the team chose this architecture two years ago. The knowledge wasn't missing — it was trapped in Slack threads, in the heads of developers who've since moved on, in docs that haven't been updated since.

That's not a documentation problem. It's a retrieval problem. And when we set out to solve this at Spotify, the results were very clear: by giving our agents real engineering context before a session even starts - same model, same codebase, same task, we reduced token cost by 32%, cut agent completion time by 65% and doubled correctness from 4/10 to 8/10. Here is how we got there.

#We built Xirp because we hit this wall ourselves

At Spotify we run one of the largest microservices environments in the world. Hundreds of engineering teams. Thousands of services. A decade of decisions baked into how everything connects.

When AI agents arrived, we felt this at scale. Agents made confident decisions that missed critical context. Engineers and agents alike burned hours orienting themselves in systems that were well-documented — just not in any way an agent could access.

So we built the infrastructure to hold all of it — services, ownership, dependencies, decisions, history — and wired it into our AI coding workflows. The context layer is Portal. The agentic development environment that runs on top of it is Xirp.

#What Xirp actually is

Xirp is a standalone desktop app built for agentic development with your real engineering context. Point it at your repos and start running agent sessions When a session begins, the agent already knows the service it's working inside, what depends on it, and what decisions have been made around it. It doesn't explore. It doesn't guess. It knows.

No platform team required. No enterprise setup. One developer can start today.

As your team grows, Portal is the natural next step — org-wide infrastructure with a full service catalog, enterprise plugins, and dedicated support. Same product, more switched on. But Xirp on its own is where it starts.

#What this means for your token spend

Every AI agent session has a hidden tax. Before the agent writes a single useful line of code, it orients — scanning the codebase, inferring service boundaries, guessing at ownership, generating something wrong, and regenerating. All of that is token spend. None of it is productive work.

Multiply that across every engineer, every session, every day — and orientation becomes one of your largest AI tooling line items. You're paying your agents to learn things your organisation already knows.

#The waste scales with adoption

When one engineer uses an agent, the orientation tax is manageable. When fifty engineers are running sessions, every one starts from zero. The agent that helped debug a service yesterday has no memory of that when a different engineer picks it up today. Same context, rediscovered and paid for, over and over.

The session without context didn't just cost more — it went down the wrong path entirely. It scored 4/10 on correctness and would have stalled before delivering anything. The session with context scored 8/10 and produced an executable plan. The expensive session wasn't just wasteful. It was wrong.

Token spend is the visible cost. Velocity loss is the hidden one.

#Context upfront changes the math

The fix isn't a cheaper model or a tighter prompt. It's giving the agent context before the session starts.

That's what Xirp does. The agent starts with your real codebase context already loaded — the service, the ownership, the dependencies, the recent decisions. It skips the exploration phase entirely.

At Spotify, we tested this. Same task, same model, same codebase — once with Workspace context from Portal, once without. Without context: $18.08, 38 minutes, unusable output. With context: $12.21, 13 minutes, production-ready plan. 32% cheaper, 65% faster, and the only session that actually worked.

The agent with context pulled the full project map — wiki pages, tracked resources, catalog entities — and used it to understand where to look before touching the codebase. One wiki page gave it a synthesized view of the requirements, architectural decisions, and how the feature connects to the broader platform. Without that, it would have missed the reasoning behind decisions that only existed in the team's documentation.

#Every session makes the next one cheaper

Xirp doesn't just consume context — it produces it. When a session ends, what was decided, built, and learned gets stored and made available. The next time anyone or any agent touches that service, that context is already there.

The more sessions your team runs, the richer the context layer becomes, and the less each subsequent session spends on orientation. Most AI tooling costs scale linearly with adoption. With Xirp, the cost curve bends.

#What this looks like for your team

One task without context: $18.08, 38 minutes, unusable output. The same task with context: $12.21, 13 minutes, production-ready plan. Multiply that across a team of 50 engineers running multiple sessions a day, and the gap becomes a significant monthly line item — in both cost and wasted time correcting wrong output.

The most expensive token is the one that teaches your agent something your organisation already knew. Xirp makes sure that token never needs to be spent.

#Start for free

Xirp is free and available now. No credit card, no procurement process, no CTO approval needed. Start with one team, one project — and see what your agents can do when they stop spending tokens on orientation and start spending them on work.