DeepSeek Super-Agent vs Claude Code: What Actually Differs

DeepSeek's Harness team has published Super-Agent, an open-source coding agent that waits for work instead of being asked for it. It landed on GitHub under the MIT license as dsh-superagent-prototype, and it is aimed squarely at the slot Claude Code occupies.

The write-up framing it calls Super-Agent "an open-source, always-on AI coding agent architecture," positioned as an autonomous alternative that shifts focus "from model capability to solving human attention limits in software development." That phrasing is the whole pitch, and it is also the main thing to check before adopting it.

Router, Workers, Watchdog

The architecture splits into three pieces. The Router is described as "an always-on, lightweight agent that monitors group chats (Slack, WeChat), GitHub pull requests, and CI/CD pipelines" — it triages and delegates, and deliberately does not execute shell commands itself. Workers are isolated sessions spun off by the Router, each using a full-capability model and shell tooling to run one task. A Watchdog runs on a timer over the active Workers, the source says, "ensuring tasks don't stall or get stuck in infinite loops."

The safety story rests on permission tiers rather than inline prompts: off, observe — where the agent logs what actions it would have taken — and auto. In observe mode a Worker prepares a full solution in a sandbox and opens a draft PR or a Slack summary for review before anything touches the base branch.

The number, and what it is not

DeepSeek reports a p95 task completion time of 15.7 minutes against 29.5 minutes for traditional single-session subagent approaches. The source labels these as synthetic benchmark tests published by DeepSeek — vendor-run, on vendor-chosen tasks, with no independent reproduction cited. Treat the roughly 2x gap as a claim about parallel dispatch, not a measured speedup on your repo.

Super-Agent vs Claude Code

The comparison in the source is less about quality than about when each one starts.

	Claude Code	Super-Agent
Interface	Terminal-first interactive CLI	Always-on background agent across Slack, GitHub, CLI
Trigger	Human @ mentions or terminal commands	Event-driven monitoring plus @ mentions
Delegation	Single session with subagent delegation	Router dispatching parallel isolated workers
License	Proprietary product / CLI	Open-source (MIT) research prototype
Best used for	Active feature engineering, interactive refactoring	Continuous triage, automated bug patching, CI repair

Two words in that table should set expectations: research prototype. A Router with read access to your Slack and your PR queue is a different security surface from a CLI you launch by hand, and observe mode is the only tier where that difference stays cheap while you find out.

The same day, a second agent release

Meituan released LongCat-2.5 on 2 November 2026, moving from the 2.5-Preview of 25 September. It is pitched at sustained work across terminals, browsers, GUIs, spreadsheets and design tools rather than at single-prompt quality. Meituan's reported scores — described by the coverage as vendor-reported and not independently verified — are Terminal Bench v2.1 at 88.4 against 70.8 for LongCat-2.0, AutomationBench 51.8, BrowseComp 91.9, and Program Bench 25.5 against a top entry of 37.

That last figure is the useful one: a model strong on terminal and browser automation and middling on program synthesis is a specialist. The coverage also flags the reported scale — 1.6 trillion total parameters, about 48 billion activated per token — as disputed, since the same figures were attached to LongCat-2.0. LongCat-2.5 is reachable through longcat.chat and the LongCat API, and the coverage says it supports Claude Code, Codex, Hermes and other agent tools.

What to watch

Both releases bet that the bottleneck is no longer model quality but how long an agent can stay on a job without a human in the loop. The honest test is cheap: point Super-Agent's observe mode at one low-traffic repo for a week and read the logs of what it would have done. If the draft PRs are ones you would have merged, the p95 number starts to matter. Until then it is DeepSeek's number, not yours.