DeepSeek has opened the Harness Developer Preview and made its code available under the MIT license, according to 36Kr’s authorized Tencent Tech report. The preview went live on the evening of August 13 Beijing time, with the company presenting an early look at a system designed for autonomous, agent-style workflows. The report adds that while the source is permissively licensed, practical operation still depends on an API key from DeepSeek, positioning the release as both open and gated in distinct ways. 36Kr’s republished Tencent Tech report provides the initial details.
Described by the South China Morning Post as a framework for turning AI models into autonomous agents, Harness is intended to let developers connect models to external software, allow them to run code, and set them up to complete complex jobs. SCMP’s overview frames Harness as a tool for creator and builder communities seeking structured autonomy patterns that can coordinate multiple steps and integrate tool use. As SCMP reports, the concept centers on wrapping models in agentic capabilities rather than keeping them as isolated predictors.
Agent framework and operational settings
According to SCMP, Harness provides developers with a framework intended to transform AI models into autonomous agents that can use external software, execute code, and complete complex jobs. The account emphasizes a design that supports multi-step execution and orchestration of tasks that require more than a single response from a model. In this architecture, the model is embedded in a broader agent loop that can interact with tools and services.
SCMP also reports that Harness offers four operational settings: standard, code-focused, creative, and minimal. The 36Kr report specifies that the code-focused mode is called PTC, short for Programmatic Tool Calling, and notes that the current build is a v0.1 candidate release with possible breaking changes. In other words, the preview signals intended modes while also indicating that interfaces, behaviors, or defaults may shift as the project evolves. The mix of modes, as described, presents different stances on how the agent should behave in typical, code-centered, exploratory, or stripped-down contexts.
The developer preview status is central to the current positioning. According to 36Kr, DeepSeek is releasing the source under an MIT license but clarifies that actually operating agents still requires a DeepSeek API key. The report further states that the preview is a candidate release, and labels it v0.1 with the potential for compatibility changes. Together these points outline early access focused on experimentation rather than a locked platform specification.
DeepSeek uses append-only session logs as the agent interaction record, according to 36Kr. This detail indicates that Harness maintains a running record of interactions as they occur, which the report characterizes as an append-only format. Within the developer preview, such a log can document agent behavior for later inspection or playback aligned with the system’s chosen operating mode.
Cordis plug-in system and modular design
A core theme in 36Kr’s report is the Cordis plug-in system that underpins Harness. According to the account, Cordis treats a wide set of elements as plug-ins that can be loaded, unloaded, and replaced. The list includes models, tools, skills, sessions, sandboxes, storage, agent loops, task scheduling, and user interfaces. That inventory spans both the computational guts of an agent and the user-facing layers, suggesting a modular construction for many moving parts in an agent workflow.
By handling these components as replaceable plug-ins, Cordis describes a model in which a developer can swap pieces in and out. The 36Kr report highlights the ability to load and unload these elements, setting the stage for developers to configure a specific blend of tooling, scheduling, storage, and interfaces matching their use cases. Within that plug-in picture, models and tools sit next to execution and presentation choices, creating a set of levers that can be adjusted inside a single framework.
Cordis’s treatment of sessions and sandboxes as plug-ins mirrors its approach to models and tools in the same report. According to 36Kr, sessions are part of the plug-in set, which connects to the note that session logs are append-only. Sandboxes and storage are listed as replaceable, placing code execution and data handling alongside the rest of the system’s modular parts. The inclusion of agent loops and task scheduling on the same footing shows that orchestration itself is subject to the plug-in approach.
For readers tracking the broader narrative around agentic AI, EastFrontier has additional reporting that provides context on related developments around the company’s direction. See recent EastFrontier coverage for more background, including our newsroom’s prior reporting on related themes in the same ecosystem.
Access, licensing, and developer preview scope
Access to the preview follows a two-part arrangement, according to 36Kr. DeepSeek opened the Harness Developer Preview on the evening of August 13 Beijing time, and the source code is released under the MIT license. At the same time, the report states that running agent operations still requires a DeepSeek API key. In practice, this combination means source availability under a permissive license with an operational dependency for live use.
The preview is identified as a v0.1 candidate with possible breaking changes, according to 36Kr. That label places the current build at an early stage, with the report signaling that interfaces or behaviors may change. It sets expectations for developers who choose to try Harness during this window, positioning the release as a work in progress rather than a finalized toolkit.
According to SCMP, Harness is intended to help developers turn AI models into autonomous agents that can use external software, run code, and complete complex jobs. This description aligns with the 36Kr focus on modular components and plug-in management through Cordis. Together, the reports present an agent framework pitched at developers who want an ecosystem for connecting models to tools within a structured workflow.
The inclusion of four operational settings, as reported by SCMP and 36Kr, gives Harness distinct run-time profiles. The standard, code-focused, creative, and minimal options provide labeled stances on how an agent session might be conducted, with 36Kr identifying the code-focused path as PTC, short for Programmatic Tool Calling. These settings live alongside the Cordis-managed components that can be loaded, unloaded, or replaced.
Session handling is defined in 36Kr’s account as append-only logs that record agent interactions. That format and the listed plug-ins together frame how a Harness session is stored and how its constituent pieces are managed. The report’s emphasis on both operational modes and component modularity outlines the primary concepts developers will encounter in the preview.
For readers following ongoing developments around infrastructure and organizational efforts connected to agentic AI work, EastFrontier continues to track relevant updates. Our broader coverage includes additional reporting from EastFrontier that situates this preview within a wider landscape.
As of this preview, the publicly described contours come from 36Kr’s authorized Tencent Tech report and SCMP’s framing. The reports collectively describe an MIT-licensed code base, a developer preview state labeled v0.1 with possible breaking changes, a Cordis plug-in system governing a broad set of components, append-only session logs, and operational modes including a code-focused PTC option. Within those bounds, Harness is presented as a framework intended to let developers transform AI models into autonomous agents that can connect to external software, execute code, and take on complex jobs.