Coding agents
Scope, DeepInfer, Humming: tooling for agents and inference goes open source
From Oct 19 to 21, Microsoft, Vercel, DeepSeek and vLLM open-sourced building blocks to test coding agents, resume them after a crash and serve models.
Between October 19 and 21, 2026, four open-source projects shipped around coding agents and inference. None of them is a new agent: they are tools to measure, harden and cheaply run the agents that already exist.
Microsoft Scope: test what agents do with your product (Oct 19)
Microsoft released Scope, an MIT-licensed evaluation platform for coding agents, built on Kubernetes (Microsoft Open Source).
- How it works: you provide a task, a starting codebase and an agent setup; Scope drives GitHub Copilot, Claude Code or VS Code with Copilot, then an LLM-based "Judge" service scores the output against criteria arranged as a graph (a runtime check can depend on a successful build).
- Traceability: tool calls, generated files and workspace snapshots are kept so a failure can be traced back to the step that caused it.
- Example given: the Azure Cosmos DB team says it resolved 20 Agent Kit issues and 27 documentation issues; after one targeted update, agents' use of hierarchical partition keys rose from 47% to 100%.
- Limits: there is no hosted version, you deploy Scope yourself, and Microsoft warns that the automated judge can make mistakes. Codex and OpenCode are "planned"; Cursor is not integrated (project site).
Vercel durable-harness: an agent that resumes after a crash
Vercel Labs released durable-harness (Apache 2.0, version 0.0.1-beta.6 of October 19 on npm), which runs Pi, OpenCode, fx, Claude Code and Codex as durable workflows (GitHub). Every model and tool call is written to the session's log: after a crash, the turn in flight resumes on another machine, and a tool call that is not safe to repeat is never run twice. Claude Code and Codex only run in the session's own Vercel Sandbox. The project is labelled experimental: its API may change.
DeepSeek DeepInfer: the production stack under Apache 2.0 (Oct 20)
DeepSeek open-sourced DeepInfer, the inference framework behind its V4.1 Flash model (GitHub). According to DeepSeek, a single cluster currently serves 155 trillion tokens per day, with overall revenue of $38 per GPU-hour at off-peak (valley) pricing (deployment Q&A).
- Hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
- What is missing: KV cache management, prefill/decode disaggregation and load balancing stay internal; the published release only does offline inference.
- See the DeepSeek V4.1-Flash-1017 tool page.
vLLM Humming: kernels for quantized models (Oct 21)
Ant Group and the vLLM community open-sourced Humming, a CUDA kernel library for quantized inference, bundled with vLLM from version 0.29.0 (vLLM blog). On DeepSeek-V4.1-Flash with eight H20 GPUs, switching the MoE layers from Marlin W4A16 to Humming W4A8 raised end-to-end throughput by roughly 14% to 43% depending on concurrency, according to vLLM's published measurements.
What it changes
- Agents become first-class users: with Scope, an SDK or MCP team can check whether agents find and correctly use its product, the way it would test documentation with humans.
- Reliability moves to the front: resuming a long task after a crash avoids paying again for model calls already made.
- Software cuts inference costs: DeepInfer and Humming chase the same goal as "Flash" models (see our article), serving more tokens on the same hardware.
Every figure in this article is announced by the vendors themselves and has not been measured independently.
Spotted an error in this article? Report it · Our corrections policy