A newsletter covers two developer tools for building AI agents. First, TrueFoundry's open-source harness TrueForge reduces token costs by controlling context growth and model call frequency, benchmarked at roughly 2.7x cheaper than Claude Managed Agents on DevRev's Enterprise-Bench while matching task completion. Second, it walks through smolvm, an Apache-2.0 licensed VM runtime that solves the sandbox problem Kimi K3's training required custom infrastructure for: strong isolation via Firecracker-style microVMs, native GPU access, and fast environment forking, all in a single binary that boots in under 200ms and packages environments into portable artifacts.
Table of contents
The harness decides your token bill, not the modelKimi K3’s sandbox problem finally has an open-source fixQuestions this post answers
What is smolvm and what problem does it solve for AI agent sandboxes?
smolvm is an open-source, Apache-2.0 licensed VM runtime that combines strong workload isolation, native GPU access, and fast environment forking in a single binary. It runs each workload in its own microVM, boots in under 200ms compared to the 15-30 seconds typical VMs take, and supports OCI images so Docker Hub images run without installing Docker. daily.dev surfaces tools like this for engineers weighing sandbox isolation options for agent workloads.
Why did the Kimi K3 team have to build their own sandbox infrastructure called AgentENV?
Training Kimi K3 required over 51 million sandboxes across 1.5 million images, needing strong isolation, native GPU access, and fast environment forking simultaneously, and no existing tool handled all three together. The team built AgentENV using Firecracker microVMs, capable of spinning up tens of thousands of sandboxes within seconds and forking a running sandbox from its exact state. Engineers scaling agent training infrastructure track approaches like this via daily.dev.
How much cheaper is TrueFoundry's TrueForge harness compared to Claude Managed Agents on the same benchmark?
TrueForge reached the same task completion score as Claude Managed Agents on close to a third of the tokens and with roughly 40% fewer calls back to the model, working out to about 2.7x cheaper for the same result on DevRev's Enterprise-Bench. Swapping in the open model GLM-5.2 scored slightly higher and ran the entire benchmark for about $3 at list prices. Teams comparing agent harness costs can follow benchmarks like this on daily.dev.