DeepSeek Harness: Free Agent Loop, Costlier Tokens 2026

DeepSeek open-sourced its agent harness, dsh, under the MIT license on August 13, 2026, and shipped the GA build of V4-Pro the same day. The harness is free and model-agnostic. The tokens it burns got more expensive three days later.

TL;DR

The runtime is now free software. The loop that runs on it is not.

Two releases on August 13 and a price change on August 16 only make sense read together:

• DeepSeek Harness (dsh) went open source under MIT, as a developer preview built on an architecture where the model adapter, tool registry, session log and the agent loop itself are all replaceable plugins.¹²
• The repository passed 161,000 stars and 16,900 forks by August 19, six days after release — up from more than 33,000 in its first few hours.¹³
• DeepSeek-V4-Pro-0813 went GA with native OpenAI Responses API support, one-click Codex setup, and three reasoning-effort levels.⁴
• New API pricing took effect at 16:00 UTC on August 16, splitting rates into peak and off-peak bands. For V4-Pro, a cached input token now costs $0.022 per million off-peak and $0.044 at peak.⁵
• The headline agent benchmarks are provider-reported. DeepSeek claims 87.9 on Terminal-Bench 2.1; the aggregator BenchLM carries that exact figure but files it as a "provider run." Artificial Analysis, which runs its own evaluations, scores the model 53 on its Intelligence Index — level with GLM-5.2 and seven points behind Kimi K3.⁶⁷⁸
• No weights have been published for the 0813 build. The deepseek-ai/DeepSeek-V4-Pro repository's last weight upload was April 23, 2026, with no commit of any kind after June 22, 2026.⁹ BenchLM states it plainly: "DeepSeek has not published weights for the 0813 hosted checkpoint."⁶

What You'll Learn

• What an agent harness is, and what "everything is a plugin" actually buys you
• The four presets dsh ships with, including the one that collapses five tool round trips into one call
• Which models and subagents the harness can drive, including Claude Code and Codex
• Exactly what the new peak/off-peak pricing costs, and why cached reads are the number agent builders should watch
• Which V4-Pro benchmark figures have been independently checked and which have not
• Where DeepSeek sits in a harness field that Chinese and US labs have both been filling since 2025

What an Agent Harness Is

An agent harness is the scaffolding around a model: the loop that reads files, calls tools, runs shell commands, and feeds each result back so the model can decide its next step. The model does the thinking. The harness does the driving, the logging, and the constraining.

That separation is the whole point. Because the harness and the model are independent, an open harness lets you swap models without rewriting your workflow — which is precisely what DeepSeek shipped.

Everything Is a Plugin

DeepSeek's stated design principle is that nothing in the system is privileged. Its architecture document is explicit: "Every part of the product is a plugin, including the model adapter, the tool registry, the session log, and the agent loop itself, so every part is replaceable from configuration."²

The same document adds that there is "no privileged core to patch" — you extend dsh by mounting a plugin beside the others, and registrations unwind when their plugin unloads.²

The architecture sits on top of Cordis, described by its authors as a meta-framework for spatiotemporal composability. Its design paper is a preprint dated August 13, 2026 — the same day the harness shipped — and is explicitly labelled as under active revision.¹⁰

Underneath it, dsh keeps an append-only session log under a rule the docs state as "Model-visible means logged": anything that reaches a model request must be reconstructable from the log, enforced by a runtime invariant.² Fork, resume, transcripts, telemetry and persistence all derive from that single stream.

That rule turns conversation history from an implementation detail into a contract, and it maps directly onto the tracing problem that makes agent runs hard to debug.

Sandboxing is a swappable backend in the same tree.² The New Stack's walkthrough describes the local implementation as per-platform: Linux Landlock via a Node addon DeepSeek wrote, macOS Seatbelt, or a Windows ACL restricted-token runner.³

The Four Presets

The New Stack's review of the release describes four presets, and the third one is the interesting one.³

Code mode changes how tools reach the model. Instead of exposing each tool as an individual function call, it hands the model an SDK and lets it write a program — so a sequence that would otherwise take five round trips runs as a single call.³

Hold onto that detail. It becomes the cheapest thing in this release once the new pricing is in front of you.

It Will Happily Drive Someone Else's Model

Nothing in the harness ties it to DeepSeek's models. The provider catalog covers Anthropic, OpenAI, AWS Bedrock, Microsoft Azure, Google's Gemini Enterprise Agent Platform and DeepSeek's own endpoint, plus custom OpenAI-compatible gateways.³

More pointed: dsh ships subagent providers that delegate work directly to Anthropic's Claude Code and OpenAI's Codex, resolving each binary from the host PATH so you supply the install and the login. Both are off by default.³

There are also bridges that run your existing hooks.json from either product against the harness's own interception points — which the README frames as a compatibility path rather than the better design.³

The harness includes an MCP client, Agent Client Protocol support, and reads both AGENTS.md and CLAUDE.md.³ If you have been following the push to standardize agent plugins and skills, this is that convergence showing up in shipped code.

One caveat on the "open source" framing. The project's own contributing guide states: "We are sorry that we cannot accept external pull requests at the moment," pointing contributors to GitHub Discussions and to building plugins instead.¹ The code is MIT. The development process is not.

The Price Change Lands on the Loop

DeepSeek's new rates took effect at 16:00 UTC on August 16 and introduced peak and off-peak bands. Off-peak is exactly half of peak. Peak hours run 01:00–04:00 and 06:00–10:00 UTC, tracking the Chinese workday.⁵

Here is the official table, per million tokens:⁵

Both models keep a 1M-token context window and 384K max output. Concurrency limits differ sharply: 2,500 for V4-Flash against 500 for V4-Pro.⁵

Two sources independently record the previous list price as $0.435 per million input, $0.87 per million output, and $0.003625 per million cached input: The Decoder's launch coverage and BenchLM's model profile, the latter dated August 17.¹¹⁶

Work the ratios against DeepSeek's published rates and the shape of the increase is clear. Off-peak input rises about 52%. Off-peak output more than doubles. Cached input rises just over sixfold off-peak — and about twelvefold at peak.

One caveat worth stating: the reseller OrcaRouter lists a different pre-change cache rate of $0.060, so that single figure is not unanimous.¹² The two sources that agree on $0.003625 also agree with each other on input and output, which is why we use them.

Why cached reads are the number to watch

What is confirmed, straight from DeepSeek's own pricing page, is the ratio: a V4-Pro cache hit costs $0.022 against $0.66 for a cache miss. That is one-thirtieth.⁵ Artificial Analysis, working from the same published rates, records it as a 97% cache discount.⁸

Agent loops live in that ratio. Every step re-sends the conversation prefix — the system prompt, the file contents already read, the tool results already returned. Those tokens are cached reads, and a long-horizon agent accumulates them faster than it generates output.

Work an example on the published numbers. An agent that re-sends a 200,000-token prefix across 50 steps moves 10 million cached input tokens in one session. At the off-peak V4-Pro cache rate that is $0.22; at peak, $0.44. If those same tokens missed cache, it would be $6.60 and $13.20.

Two levers fall out of that arithmetic, and both are free. Schedule long agent runs outside 01:00–04:00 and 06:00–10:00 UTC to halve the bill. And cut the number of round trips.

Which is where Code mode stops being a curiosity. Collapsing five tool calls into one program removes four re-sends of the whole prefix. The feature that looks like a developer-experience nicety is a direct answer to the pricing change that landed three days after it.

If you are instrumenting this properly, it belongs alongside per-session spend caps for agents rather than in a monthly invoice review.

The Benchmarks Are DeepSeek's Own

The agentic gains from preview to 0813 are large, and every one of them comes from DeepSeek. The rows below are the ones two separate aggregators both carry, each tracing the figure back to DeepSeek's own announcement and technical report.¹²⁶

DeepSeek published more figures than these; we list the ones we could corroborate in two places. All are provider-reported.¹²⁶

The label matters more than any single number. BenchLM carries DeepSeek's 87.9 under the name "Terminal-Bench 2.1 (provider run)" with an evidence status of "Provider exact." It reserves a separate status, "Benchmark exact," for rows an outside evaluator produced — the Vals AI Vibe Code Bench score of 49.93, for instance.⁶

The independent picture is cooler, and it comes from evaluators who ran the model themselves. Artificial Analysis scores it 53 on its Intelligence Index. As the South China Morning Post reported the day of release, that puts it level with Zhipu AI's GLM-5.2, four points behind GPT-5.6 Terra and seven behind Moonshot's Kimi K3.⁷⁸

On the Vals Index it ranked 12th, trailing the previous-generation GPT-5.5. Vals AI named two specific weak spots: completing tasks within a sandboxed terminal environment, and generating complex financial models in Excel spreadsheets.⁷

There is one more datapoint, and it needs its provenance stated exactly. OrcaRouter's model page lists two Terminal-Bench 2.1 entries for this model — DeepSeek's 87.9, and a second reading of 78.7 filed under the raw benchmark slug. The page credits its benchmark data to Artificial Analysis and DeepSeek's API docs, and Terminal-Bench v2.1 is indeed one of the nine evaluations Artificial Analysis runs itself.¹²⁸ We could not confirm the 78.7 on a second site, so treat it as suggestive rather than settled.

This is the harness confound in its purest form. Agentic scores are a property of the model and the scaffolding it runs inside.

The harness repo's own BENCHMARK.md documents running benchmarks through the jsonrpc-agent minimal variant, and The Decoder reports DeepSeek uses that stripped-down minimal setup for its own benchmark runs.¹¹¹³ Open-sourcing the harness makes that scaffolding inspectable, which is a real contribution. It does not make the numbers reproduced.

The weights question, where sources disagree

The 0813 build's weights have not been published. Our own check of the deepseek-ai/DeepSeek-V4-Pro repository on August 19, 2026 found the last weight upload dated April 23, 2026, and no commit of any kind after June 22.⁹ BenchLM says the same in one line — "DeepSeek has not published weights for the 0813 hosted checkpoint" — and lists the model's self-host status as "Weights are not published."⁶ The Decoder reported it on release day too.¹¹

Artificial Analysis disagrees, labelling the model "open weights" under an MIT license and linking to Hugging Face for the download.⁸ That link points at the same repository we inspected — the one still carrying the April build. The most likely reading is that the open-weights label attaches to the V4-Pro family rather than the August checkpoint being served by the API.

Either way, if you plan to self-host the model behind those 87.9s, check what you are actually downloading. That is a different posture from DeepSeek's open-weight V4 release in April, where the weights and the announcement arrived together.

DeepSeek Is Not First Here

The harness field was already crowded. Among Chinese labs alone, Alibaba's Qwen Code and ByteDance's Trae Agent both launched in mid-2025, Moonshot's Kimi CLI followed that October, and ZCode from Zhipu AI — now trading as Z.ai — arrived in July 2026.³¹⁴

What makes the timing land is the direction of travel elsewhere. Google announced in May 2026 that Gemini CLI would stop serving Google AI Pro, Ultra and free Code Assist users on June 18, pushing them to Antigravity CLI — whose public repository, when The Register checked it in May, held only a changelog, a readme and a GIF.¹⁵ The Gemini CLI repository itself stays up under Apache 2.0, with Google continuing to maintain it for enterprise customers who still have access.¹⁵

So the honest framing is not that DeepSeek invented open agent tooling. It is that DeepSeek opened a harness roughly two months after a major US lab replaced its open CLI with a closed-source successor, and developers noticed at a scale — 33,000 stars in hours, past 161,000 in six days — that is hard to read as anything other than appetite for a runtime nobody can take away.¹³

What This Means If You Build Agents

Four practical reads.

The harness is worth evaluating on its merits, not its origin. It drives Anthropic, OpenAI, Bedrock, Azure and Google endpoints, and its append-only session log gives you a reconstructable record of everything the model saw. Nothing forces you to point it at DeepSeek.

Treat the developer-preview warning literally. The README warns of compatibility-breaking changes; the contributing guide says external pull requests are not being accepted.¹ That is fine for evaluation and risky for a production dependency this quarter.

Model your token bill on cached reads, not output. The 30× gap between a cache miss and a cache hit, and the 2× gap between peak and off-peak, are the two levers that actually move an agent's cost — and both are published.⁵

Do not spend production budget on DeepSWE 62.7 yet. A jump from 12.8 to 62.7 in one release is a provider claim, filed by aggregators as exactly that. The measured picture — Artificial Analysis 53, Vals 12th — is the one you can act on today.⁶⁷⁸

Bottom Line

DeepSeek gave away the part that used to be a moat and started charging more for the part that was already commoditized.

The harness is a serious piece of engineering with an unusually clean plugin story, a strong session-log guarantee, and no lock-in to its author's models. It deserves the attention it got.

The model that shipped alongside it is a mid-pack frontier model on measured indices with a vendor scorecard that outruns its independent record, running on tokens that repriced upward three days after it shipped. Both statements are true. Read the release as two announcements, not one.

Sources