Alchymos x G2i

AI teams are learning that the cost of an agent is shaped by more than the model selected at the start of a request. It is also shaped by how much context the system sends, how often tools are called, how many times a workflow retries, and whether engineers can see spending early enough to act.

At G2i, we saw that challenge firsthand while building infrastructure for evaluating advanced AI agents against realistic applications. The work involved complex, multistep workloads across APIs, Model Context Protocol tools, and user interfaces. Those evaluations needed to be repeatable, observable, and scalable, and they could also generate a substantial amount of model usage.

The opportunity: make every token work harder

The goal was not to make agents less capable. It was to remove avoidable work around them. We looked at the full evaluation workflow and focused on a simple question: which requests truly need an expensive model call, and which ones can be handled with better context management, cheaper defaults, or a result that has already been computed?

That led to a set of practical optimization techniques inspired by the same principles behind Alchymos: cache what is safe to reuse, keep context focused, and make the cost of agent behavior visible enough that teams can improve it continuously.

A high-level approach to optimization

We introduced several layers of discipline across the workflow:

• Better context management: reduce unnecessary material entering a request and keep the agent focused on the task at hand.
• Model routing by task: default to less expensive models for straightforward work, reserving the most capable models for tasks that genuinely require them.
• Smarter tool boundaries: avoid expensive MCP calls when they are not needed and prevent workflows from repeatedly processing generated files or dependency directories.
• Lean internal guidance: keep skills and instructions useful and focused instead of sending large amounts of redundant context.
• Clearer spend visibility: give engineers earlier signals through smaller spending thresholds and more immediate alerts.
• Helpful nudges: encourage the team to match the model to the task, rather than using the most expensive option by default.

None of these ideas depends on a single optimization. Together, they create a feedback loop: understand where tokens go, remove waste, measure the effect, and make the better path the default.

The impact

Using these techniques across the AI workflow, G2i reduced recurring AI expenditure by approximately $60,000 per month. The result came from improving the system around the models, not from asking engineers to do less or compromising the quality of the evaluation work.

The most important lesson was that cost optimization is an architectural concern. It belongs in the same conversation as reliability, observability, and developer experience. When teams can see expensive patterns and have practical controls for changing them, meaningful savings can come from many small decisions made consistently.

Why this matters for agent builders

As agents move from demos into production, model spend becomes a property of the whole workflow. A single request may fan out into multiple model calls, tool calls, retries, and validation steps. Without visibility, it is easy to pay repeatedly for work that adds little value.

The G2i experience reinforces the case for agent-native infrastructure: caching, context controls, model-aware routing, execution traces, and cost alerts should work together. The right optimization is not always a cache hit. Sometimes it is a smaller prompt, a cheaper model, a skipped tool call, or an engineer seeing a warning before a pattern becomes a monthly bill.

Alchymos is being built around that perspective. We help teams make repeated work reusable, agent execution understandable, and optimization an ongoing part of operating AI systems.

If you are working through similar model-cost challenges, try Alchymos or book a conversation.