Announcement
Introducing Eragon Flux
Model routing for enterprise agents, built into the Eragon harness.
Over the past year, we've watched AI usage at our customers spread from a few teams to entire companies. Agents now prepare account reviews, reconcile ledgers, triage tickets and ship code every day. As usage has grown, so has the share of budget going to tokens, and much of that spend goes to work that doesn't need the most capable model.
Today we're launching Eragon Flux, model routing built into the Eragon harness. Flux breaks each task into steps and assigns each step a model. We reserve the most capable models for the hardest steps and run the rest on the best model available at the lowest price. Flux keeps evaluating models against your real work, so cost per task continues to fall after you turn it on.
Flux is already in production at Deel, Slash, Prodware and many other Eragon customers.
Most agents run every step on the most expensive model
A single request is usually many smaller tasks. To prepare a quarterly business review, an agent pulls the account from the CRM, reads recent support tickets, summarizes call notes, checks product usage and then writes the narrative.
Writing the narrative takes judgment. Most of the other steps are retrieval, extraction and formatting, which a smaller model can handle at a fraction of the cost and often with lower latency. Most agents send all of it to the same frontier model anyway.
In a pilot, that overhead is easy to miss. Once a workflow runs across every account each week, it becomes one of the largest lines in the AI budget, and often the reason the workflow stops expanding.
Model selection shouldn't fall to the people doing the work
Someone closing the books or fixing a bug is focused on the work in front of them. Deciding which model should handle each step isn't part of that job, and it shouldn't have to be.
It's also a hard decision to keep making well. New models ship every few weeks, and prices and quality change with them. Without time to evaluate each one, most teams default to the most capable model for everything. Flux makes that decision for them, step by step.
Flux plans each task and routes every step to the right model
When a task comes in, a planning agent breaks it into steps and decides what each one requires: how much reasoning, which context and tools, and what kind of output. It then assigns each step to a specialized subagent running on the model best suited to it.
- The hardest steps go to the most capable models. These are the judgment calls, complex reasoning and final synthesis that decide the quality of the result.
- Everything else goes to the best-value model. Flux picks the model that handles the step well at the lowest available price.
- Every step is verified. Outputs are checked before the next step starts, and hard or low-confidence steps are rerun on a more capable model.
Every subagent works from the same Eragon context, including company knowledge, memory and permissions. Because Flux is model agnostic, it routes across open and closed providers and chooses on quality and price alone.
Smaller, focused models are often more accurate
Routing doesn't have to trade quality for cost, and in our deployments it often improves quality. A subagent that sees only the context its step needs is less likely to be thrown off by irrelevant information. A narrowly scoped step is easier to get right than a long, open-ended prompt. Verifying each step catches errors before they carry into the rest of the task.
Flux gets cheaper over time
Flux learns from every task it routes, building a picture of which models perform well on which kinds of steps for each customer's workloads. It runs evaluations of cheaper models against those real tasks, and moves traffic to a different model only when quality holds. Teams don't need to run their own benchmarks or rewrite prompts to benefit.
Flux works in your coding tools and across knowledge work
Coding. Create an Eragon key through the Eragon AI gateway. The key carries your Eragon context and works with any coding client, including Claude Code, Cursor and Codex. Engineers keep the tools they already use, and Eragon routes each request behind the scenes.
Knowledge work. Flux automatically runs in the Eragon platform client across GTM, finance and operations workflows. In account research, a top-tier model chooses the angle and writes the outreach, while lower-cost models gather CRM history and recent news. At month-end close, a top-tier model investigates exceptions, while lower-cost models match transactions and draft reconciliations.
Lower costs let teams automate more
Much of the work companies want to automate happens at high volume, across thousands of accounts, invoices or tickets. At that scale, cost per task decides which workflows make it to production.
We built Flux so teams that have already scaled AI, or want to significantly scale AI from the start, can keep expanding without costs rising at the same rate. Saving the most capable models for the hardest work makes it economical to automate far more of what a company does each day. That is the future of tokens compounding that we're building towards at Eragon.
Get started
Eragon Flux is available today. Reach out at hello@eragon.ai to roll it out across your team.