The agent stack DeepSeek had published by November 2, 2026 is a MIT-licensed, plugin-based agent harness — the software layer that wraps a model with tools, files, and permissions so it can carry out multi-step work — plus an OpenAI- and Anthropic-compatible model API with priced web Search and Fetch tools. An enterprise SaaS team can build agent features on that stack today, on two conditions: pin the harness against compatibility-breaking changes, and budget for tool-call billing that, under DeepSeek's published terms, began on November 1.
The trigger is a headline that outran the record. On November 2, an item reading "DeepSeek亮出'超级智能体'架构" ("DeepSeek unveils a 'super agent' architecture") sat at No. 20 on the Toutiao hot board, the trending list of ByteDance's Chinese news aggregator.[1] No official release, repository, or documentation page we could find carries that name: DeepSeek's news index and change log both end at the October 17 release, and its published agent-harness repositories contain no such product.[4][5][2] What DeepSeek did publish in October is narrower than the headline, documented first-party, and relevant to SaaS integration work.
The official record: an agent harness, not a "SuperAgent"
The core artifact is DeepSeek Harness, an open-source agent harness that the company describes as built on an "everything-is-a-plugin" architecture powered by Cordis.[2] The architecture documentation is specific about what that means: every part of the product is a plugin — the model adapter, the tool registry, the session log, the agent loop itself — so each is replaceable from configuration, and extension happens by mounting a plugin beside the others rather than patching a privileged core.[3] The project is MIT-licensed, installs with npx @deepseek-ai/dsh web, and serves its Web UI at 127.0.0.1:3080 by default.[2] It is also in developer preview, and the README states that there will be compatibility-breaking changes.[2] DeepSeek's own API documentation lists the Harness first under its Agent Integrations section, with an official setup guide.[9]
Around the harness sits the model layer. The deepseek-flash API model resolves to DeepSeek-V4.1-Flash-1017, released October 17 and live on the company's app, website, and API, with open weights on Hugging Face and ModelScope; deepseek-v4-pro resolves to DeepSeek-V4-Pro-0813.[4][7] Both models take 1M-token contexts and produce up to 384K output tokens.[7] The third piece is the tool layer: Search and Fetch APIs, which DeepSeek opened to all API users on the same October 17 date at a published $0.75 per 1,000 calls per endpoint.[4]
What the plugin architecture means for a SaaS integration
The integration path starts with compatibility, and this is where the published record is concrete. The API exposes two base URLs — https://api.deepseek.com in OpenAI format and https://api.deepseek.com/anthropic in Anthropic format — and both models support Tool Calls, the Responses API, and the Anthropic API.[9][7] Agent code already written against either format therefore changes a base URL and an API key, not its calling conventions.
The tool endpoints follow the same pattern. Search and Fetch are raw JSON APIs at POST /v1/tools/search and POST /v1/tools/fetch, using the same bearer authentication as the model API; Search returns structured results with a count parameter defaulting to 10 and capped at 50, Fetch returns a page as Markdown, and Fetch carries a per-domain QPS limit (queries per second) on top of account-level limits.[6] For a SaaS feature that needs retrieval — support-ticket triage over documentation, competitive monitoring, retrieval-augmented generation (RAG) over live web content — a backend calls these endpoints directly with published schemas and no SDK dependency.
Teams that want the harness itself get a defined extension surface. The shipped profiles are web, headless, sdk, sdk-minimal, and acp, and the dsh-base bundle — the shared first layer of all of them except sdk-minimal — ships model adapters, tools, persistence, sandbox and approval policy, settings, credentials, and telemetry.[3] The Web UI agent can read and edit workspace files, run commands, delegate work, and maintain a plan, asking before operations the permission policy gates.[10] Extension points documented in the architecture include same-session objectives through ctx.goals, forking a session at a turn boundary, and swapping the session-persistence backend.[3]
The breadth of existing integrations lowers the trial cost further. DeepSeek's official Awesome DeepSeek Agent repository curates 23 integration guides, including Claude Code, Codex, GitHub Copilot, Cline, Cherry Studio, and Alibaba's Qwen Code; the API docs carry first-party setup guides for eight of them.[8][9]
What it costs, and what can break
Published prices set the budget. For deepseek-flash, 1M input tokens cost $0.15 off-peak and $0.30 at peak on a cache miss — $0.003 and $0.006 on a cache hit — and 1M output tokens cost $0.60 off-peak, $1.20 at peak. deepseek-v4-pro runs $0.66/$1.32 on cache-miss input and $1.98/$3.96 on output. Off-peak rates are half of peak, with peak hours defined as 01:00–04:00 and 06:00–10:00 UTC Monday through Friday, excluding Chinese public holidays; concurrency caps sit at 2,500 for flash and 500 for pro.[7]
Tool calls carry their own line item: $0.75 per 1,000 successful requests, per endpoint, on both Search and Fetch.[7] The published terms state that server-side tool calls made through DeepSeek Harness, Claude Code, or Codex are currently free, with per-request billing starting at 00:00 Beijing time on November 1, 2026 — the same moment the October 17 news post gives as October 31, 16:00 UTC.[7][4] As of November 2 the pricing page still carried the "currently free" wording, so the date to plan against is the published November 1 start.[7] At published rates, an agent feature making 10,000 tool calls a day spends about $7.50 a day on tools, before token costs.
Two caveats bound the recommendation. The harness is a developer preview whose README promises compatibility-breaking changes, and DeepSeek's own release note recommends Harness v0.2.1-rc.1 or later for the best experience with the new model — version churn is a stated property, not a risk to discover.[2][4] And the benchmark figures attached to the October 17 model, such as DeepSWE v1.1 at 77.0 and Terminal-Bench 4.0 at 42.3, are DeepSeek's own published numbers, tested on named harnesses; they are the vendor's claim, not an independent measurement.[5]
Bottom line for enterprise SaaS builders
The API layer is the part to adopt: both formats are compatible, the tool endpoints are documented, and every price is published, with the November 1 billing date now in force under the published terms.[9][7] The Harness is the part to pin: the plugin architecture replaces components without patching a core, which suits teams that need to swap model adapters or persistence, but developer-preview status means locking versions and reviewing the project's safety notice before any production dependency.[2][3] The headline's "super agent architecture" is the part with nothing to evaluate: no official release, repository, or documentation page we could find carries that name as of November 2.[4][5][2] The record, checked against DeepSeek's own channels, supports building on what the company published — and waiting on the architecture it has not.