DeepSeek Harness v0.1.0-rc.8: Turning Claude Code and Codex into Sub-Agents

On August 19, 2026, DeepSeek dropped Harness v0.1.0-rc.8 — just two days after rc.7 and only six days after the framework’s MIT-licensed debut. The release cadence alone is a statement: Harness is not a side project. It is the engineering shell DeepSeek wants wrapped around every serious agent workflow.

This update brings fourteen changes, but three stand out because they reshape how developers will compose AI labor:

1. Native image input is now supported across core commands like /goal and /plan.
2. Claude Code and Codex can be installed as sub-agent bundles and orchestrated from inside Harness.
3. web_search runs concurrent queries, cutting the latency of multi-question research.

Together, these features point to a larger ambition: DeepSeek is not merely building another coding agent. It is trying to own the scheduling layer of the agent era.

What Harness Actually Is

The official formula is short and precise: Agent = Model + Harness.

If the model is the brain, Harness is everything else — file access, tool use, memory, sandboxing, error retries, task decomposition, and result delivery. The framework is built on Cordis, a microkernel where everything is a plugin: the model adapter, tools, skills, sessions, storage, the agent loop, and even the UI. Nothing is welded shut.

That plugin architecture is what makes the rc.8 sub-agent feature possible. Because Harness treats external agents the same way it treats any other tool, Claude Code and Codex become composable components rather than competing products.

The Sub-Agent Play: Competitors as Components

In rc.8, Claude Code and Codex ship as Profile Bundles. Install the bundle, and either agent can be invoked as a task executor inside a Harness workflow. Codex adds non-interactive permission mode and supports multiple named instances, so a single parent task can spin up several Codex workers in parallel, each handling a different file or module.

A reportDelivery mechanism wakes the parent task as soon as a sub-agent finishes, eliminating the blocking wait that kills multi-agent efficiency.

The strategic message is hard to miss: model loyalty is fragile, but workflow loyalty is sticky. By letting users keep their existing Claude Code or Codex investment while orchestrating it through Harness, DeepSeek turns rival products into peripherals.

Native Vision and the “Folk Vision” Fallback

rc.8 also adds native image requests for DeepSeek model adapters. Screenshots can now be dropped directly into /goal, /plan, and related commands, with the @ menu gaining file and history references.

The more clever detail is the fallback path. When a model cannot accept images, Harness does not give up. It calls OCR, color statistics, pixel scanning, and other tools to convert the image into structured text. The result is a kind of “folk vision” — a pure-text model seeing through tool orchestration.

Concurrent Search and Tool Chain Speed

Multi-question research gets faster because web_search now fires queries in parallel. Windows users get a persistent PowerShell PTY session, and SQLite-backed large history sessions are faster after a backend optimization round.

One practical warning: rc.8 changes the SQLite schema. If you have been running Harness locally since the first preview, back up your data before upgrading.

Why the Scheduling Layer Matters

The AI industry has spent 2026 obsessed with model benchmarks. DeepSeek’s Harness bet is that the next moat lies one layer above the model: the runtime that decides which model handles which task, which tool to call, when to delegate to a sub-agent, and how to recover when something breaks.

Closed products like Claude Code and Codex tie you to their models and their UI. Harness is open-source, model-agnostic, offline-capable, and cheap to run on DeepSeek V4-Flash. The Android analogy is already circulating: DeepSeek does not need to make every phone; it needs every phone to run its system.

What This Means for DeepThink Users

DeepThink’s reasoning engine is the production core inside DeepSeek V4 Pro and V4 Flash. When Harness schedules a sub-agent or a long-horizon tool loop, it is DeepThink-style reasoning — transparent, self-verifying, tool-aware — that keeps the plan coherent.

For developers, the rc.8 release means you can now build agent teams where DeepThink handles the strategy, Claude Code handles a front-end refactor, Codex handles a backend migration, and Harness coordinates the handoffs. The era of single-model, single-agent workflows is ending.

Looking Ahead

DeepSeek has already trademarked the “DeepSeek Harness” name and published brand guidelines. The legal housekeeping signals ecosystem intent: open source the runtime, own the brand, and let the community build the plugins.

If the weekly release rhythm continues, Harness could move from developer preview to production-ready faster than any comparable framework in the market. And if the sub-agent pattern expands to more agents — GitHub Copilot, Cursor, Kimi K3, and open-source alternatives — Harness could become the default router for compound AI systems.

The model war is far from over. But with rc.8, DeepSeek makes a credible claim that the next battle will be fought over who controls the agent layer — not just who trains the biggest model.