Agent harnesses are moving from an experimental layer to a production product
Microsoft made its agent harness generally available while DeepSeek invited open-source developers into beta testing, placing the two companies at different stages of productisation.
Artificial Intelligence··Evening
Microsoft’s supported runtime
Microsoft’s Agent Framework Harness and Foundry Hosted Agents have reached general availability. Framework 1.0 became generally available on April 2, while the harness and the GitHub Copilot SDK and Claude Agent SDK connectors reached stable release at Build 2026 on June 2–3. The latest step turns the runtime around a model into a supported product. By default, the harness includes function invocation, per-call history persistence, context compaction, a todo list with planning and execution modes, file memory, skills, web search, tool approval, and OpenTelemetry traces; each component can be removed separately. Shell tools, file access, background sub-agents, and automatic looping are optional and issue warnings when enabled. The connectors let GitHub Copilot and Claude agents operate within existing identity, content-safety, and observability policies. Offered for .NET and Python, the framework supports sequential, parallel, and Magentic orchestration patterns, while Foundry Hosted Agents are billed according to consumption.[1]
DeepSeek’s open-ended beta invitation
DeepSeek is entering the same product area from an earlier stage. In a social-media announcement on Saturday, the company invited open-source project developers to apply for beta testing of DeepSeek Harness. The software is intended to turn DeepSeek language models into agents that can run multi-step code, reason through complex workflows, and act autonomously. According to the South China Morning Post, the team is led by Cui Tianyi, who was hired in March after working as a software engineer at Jane Street and co-founding TSY Capital. The announcement gave no release date. The newspaper reports that the team is short-staffed and recruiting through several channels. The invitation follows the V4-Flash model, but it was not accompanied by a repository, benchmark result, or licence terms. What developers can access is therefore an invitation to participate in work on the layer, rather than a downloadable public release with settled conditions.[2]
A separate product layer around the model
Both developments make visible, at product level, that an agent is more than a language model. Sustaining a multi-step task requires tool calls, state and history management, permissions, loop limits, orchestration, and observability to be assembled in a separate runtime. Microsoft has reached general availability after specifying which components are included, which remain optional, which connectors inherit organisational policies, and how hosted use is billed. DeepSeek is still seeking developer participation, with the team, repository, licence, and release timetable not yet fully presented. The term harness therefore describes two different access states: one is a runtime available for .NET and Python with documented features and a hosted option; the other is a development programme recruiting testers from the open-source community. The current movement in agent products is consequently not only another model announcement. It is the layer that governs how models use tools and enter organisational systems becoming a product category in its own right.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
