Fireworks Expands into Sandboxed Agents

The important shift is that agent infrastructure is collapsing two budgets, tokens and compute, into one platform decision. Once a customer wants the same system to call a model, run generated Python, open a browser, save files, and spin up many parallel jobs safely, Fireworks stops competing only with inference APIs and starts competing with the serverless GPU and code execution stack offered by Modal, Baseten, and RunPod.1 2 3 4 5 6

• Fireworks comes from model serving first. Its core product is low latency inference, tuning, and open model deployment, and it reached $1B annualized revenue by June 2026. Modal came from cloud code execution first, then used the same container and autoscaling base to push into agent sandboxes and also reached about $1B annualized revenue by October 2026.1 2 3
• Baseten and RunPod sit between those poles. Baseten is turning model hosting into a broader agent runtime with hosted tools for web search, browser use, and sandboxed code execution. RunPod already sells elastic GPU capacity for deploying and training AI workloads, so adding isolated agent execution makes it a closer substitute for teams that mainly care about cheap scalable compute.4 5 6
• E2B and Daytona attack the same workflow from the opposite end. They start with the isolated computer itself, a disposable environment where an agent can run commands, edit files, and use a browser, then move outward toward the inference and orchestration layers around it. That makes sandbox quality, startup speed, persistence, and security part of the buying decision, not just model quality or token price.7 8 3

This market is heading toward bundled agent clouds where inference, sandboxing, storage, and observability are sold together. The winner is likely to be the platform that makes an agent feel reliable in production, not just fast on a benchmark, because customers will pay for fewer failed runs, safer execution, and less glue code across separate vendors.2 3 4 7 8