DeepSeek Harness RC.8 Adds Multimodal Input and Agent Orchestration
DeepSeek has released DeepSeek Harness RC.8, only two days after RC.7, expanding its open-source Agent framework with stronger multimodal support, deeper sub-agent orchestration, improved tool execution, Windows terminal capabilities, and SDK updates.
The release is available through the project’s GitHub repository.
While RC.8 contains a broad collection of improvements, two changes stand out: multimodal Agent workflows are becoming more complete, and external coding Agents such as Claude Code and Codex can now operate as sub-agents inside Harness.
That combination points toward a broader architectural direction for DeepSeek Harness: rather than functioning as a single coding Agent, Harness is increasingly positioned as an orchestration layer capable of coordinating multiple specialized Agents and tools.
🖼️ Multimodal Workflows Become More Capable #
RC.8 significantly expands how DeepSeek Harness handles images and mixed-media context.
The DeepSeek model adapter can now directly issue native image requests, while core commands such as /goal and /plan support combined text-and-image input.
This allows users to provide visual information directly as part of an Agent workflow rather than converting everything into text first.
Potential workflows include:
- Analyzing screenshots during software development
- Inspecting UI designs and implementation differences
- Reviewing diagrams alongside textual requirements
- Passing visual debugging information to an Agent
- Combining source files with screenshots or other visual references
File and session references also expand #
The @ menu now provides references for files and session history.
This gives Agents more direct access to local project artifacts and previous conversational context, making it easier to construct tasks from existing work rather than repeatedly copying information into a prompt.
Combined with native image support, the result is a richer context pipeline in which an Agent can work with text, images, files, and previous sessions as part of the same workflow.
🤖 Claude Code and Codex Become Harness Sub-Agents #
The most strategically significant change in RC.8 may be its deeper integration with Claude Code and Codex.
Both coding Agents can now be installed on demand through Profile Bundles and invoked by DeepSeek Harness as sub-agents.
This changes the role of Harness considerably.
Instead of requiring users to select a single coding Agent for an entire task, Harness can act as the higher-level coordinator while specialized Agents perform individual pieces of work.
Codex gains parallel instance support #
RC.8 also adds a non-interactive permission mode for Codex.
This is particularly useful for automated Agent workflows because sub-agents can perform tasks without requiring continuous interactive approval.
Codex can additionally run multiple named instances simultaneously, making it possible to organize several independent Agent tasks within the same orchestration environment.
For larger development workflows, this creates the foundation for parallel execution rather than strictly sequential Agent interaction.
🧩 RC.7 Established the Sub-Agent Foundation #
The changes in RC.8 build directly on the architecture introduced in RC.7.
Released on August 17, RC.7 connected Codex and Claude Code sub-agent tasks to the Job Panel, allowing users to monitor and manage their execution directly from Harness.
RC.8 takes the next step by turning the corresponding Claude Code and Codex providers into independently installable Profile Bundles.
This approach gives Harness a more modular architecture:
- Harness decomposes the overall task.
- Individual jobs are created for specialized work.
- Appropriate coding Agents are selected as sub-agents.
- Sub-agents execute their assigned tasks.
- Results are returned to the parent workflow.
- Harness aggregates the outputs into the larger task.
The architecture is increasingly similar to a multi-Agent runtime rather than a conventional coding assistant.
🔎 Tool Calls Gain Parallel Execution #
RC.8 also improves the efficiency of tool-driven workflows.
The web_search tool now supports concurrent queries, allowing multiple searches to execute in parallel instead of waiting for each request to finish sequentially.
This matters particularly for research-heavy Agent tasks.
For example, a sub-agent may need to gather information from several independent sources before returning its findings to the main Agent. Concurrent execution can reduce the latency of this research stage and improve overall workflow throughput.
reportDelivery Enables Faster Parent-Task Resumption
#
Another important change affects communication between sub-agents and their parent tasks.
After a sub-agent finishes its work, reportDelivery can return the result promptly and actively wake the parent task.
This reduces unnecessary idle time in long-running Agent chains.
The resulting execution model is closer to an asynchronous task system:
- Parent Agent assigns work
- Sub-agent executes independently
- Tools run concurrently where possible
- Results are delivered immediately
- Parent workflow resumes when the required information becomes available
For complex Agent orchestration, these seemingly small scheduling improvements can have a meaningful impact on end-to-end latency.
🪟 Windows Support Gets Practical Improvements #
Windows users also receive several useful updates in RC.8.
The PTY terminal now supports persistent PowerShell sessions, and the capability is enabled by default in the Minimal preset.
Persistent terminal sessions are important for development workflows because they preserve shell state between commands, reducing the overhead of repeatedly recreating environments.
The release also simplifies installation and startup workflows.
Other changes include:
- Reduced dependency download sizes
- Improved local launch behavior
- Automatic browser opening when running
dsh web - Better Windows-oriented terminal integration
These changes make Harness more practical for developers who use Windows as their primary development environment.
🛠️ RC.8 Fixes Several Real-World Workflow Issues #
Beyond new features, RC.8 addresses a number of issues encountered during actual Agent usage.
Image payload handling #
The update fixes model request failures caused by:
- Oversized individual images
- Excessive cumulative image history
- Large visual payloads accumulated across conversations
This is particularly relevant now that multimodal workflows are becoming a more central part of Harness.
Streaming and forked-session behavior #
RC.8 also fixes an issue where a displayed response prefix was not preserved when users canceled streaming generation and subsequently continued with new questions or forked sessions.
Maintaining conversational state correctly is important for long-running Agent workflows where sessions are frequently interrupted and resumed.
OpenAI-compatible gateway compatibility #
The release additionally addresses compatibility problems with certain custom OpenAI-compatible gateways.
Some gateways previously failed because of differences in request formats or because returned reasoning content was not properly handled.
These fixes should make Harness more tolerant of heterogeneous model-serving infrastructure.
🐍 Python SDK Expands Built-In Agent Support #
The Python SDK also receives dependency improvements.
Its configuration now covers four built-in Agent presets, including dependencies required for:
rgandglobsearches- MCP stdio tools
- Built-in Agent execution environments
This helps simplify environment setup for developers building workflows around the SDK.
Rather than manually installing every dependency required by individual Agent capabilities, the SDK can provide a more complete predefined environment.
🏗️ DeepSeek Harness Is Becoming an Agent Orchestration Layer #
The most important aspect of RC.8 may not be any individual feature.
Taken together, the changes reveal a clear architectural direction.
DeepSeek Harness is increasingly separating task orchestration from task execution.
Harness can operate at the upper layer, decomposing objectives and coordinating workflows, while specialized coding Agents such as Claude Code and Codex handle specific implementation tasks underneath.
That architecture offers several advantages.
Specialized Agents can be selected per task #
Different coding Agents can be assigned to different workloads depending on their capabilities.
A parent workflow could potentially delegate one task to Codex, another to Claude Code, and use additional tools or Agents for research, testing, or validation.
Parallel execution becomes more practical #
Concurrent tool calls and multiple named Agent instances allow independent tasks to execute simultaneously.
This is particularly valuable for large software-engineering workflows where research, code analysis, testing, and implementation do not necessarily need to occur sequentially.
The orchestration layer becomes the primary abstraction #
The user no longer needs to think exclusively in terms of which Agent should perform an entire project.
Instead, the higher-level question becomes how the overall workflow should be decomposed and which specialized Agents should execute each component.
That is a fundamentally different model from a traditional single-Agent coding assistant.
🚀 RC.8 Signals a Broader Multi-Agent Strategy #
DeepSeek Harness RC.8 represents a meaningful step toward a unified multi-Agent development environment.
The release combines native multimodal input, richer context references, modular Agent providers, Claude Code and Codex sub-agents, concurrent tool execution, asynchronous result delivery, and improved platform support.
The most significant development is therefore architectural rather than cosmetic.
DeepSeek Harness is increasingly turning external coding Agents into interchangeable teammates within a larger orchestration system.
If this direction continues, Harness could evolve from an open-source Agent framework into a general-purpose control plane for coordinating coding Agents, tools, models, and long-running development workflows.
RC.8 may ultimately be remembered less for adding any single feature and more for making that multi-Agent architecture substantially more practical.