DeepSeek Super-Agent

Open-sourced by the DeepSeek Harness team in October 2026 under MIT. The difference from a normal coding assistant: it notices problems nobody asked it to look at.

What it does

Property	Detail
Scope	One GitHub repository per agent
Trigger	Watches GitHub; can also connect to Feishu or similar chat
Flow	Authorised person requests work → agent opens a work item → works in its own session → delivers back in the original thread (e.g. a pull request)
Oversight	A web panel lists items that need a human
Licence	MIT

Three accept modes

Mode	Behaviour	Good for
off	Inferred work is never even listed as a candidate	Conservative teams, first trial
confirm	Proposes candidates, waits for approval	Most teams start here
auto	Acts on its own when team-written rules allow	Mature processes with clear rules

The default is conservative: the agent records work it would accept but does not act. Reaching auto mode requires the team to have reviewed decisions and written the rules first.

The design principle is the interesting part

The project states its goal as: occupy as little human attention as possible, create as much value as possible. An agent that can work for a whole team can also waste a whole team's attention — so it speaks only when asked, or when rules permit, and only when someone must act or can use the result.

How it changes your bill

	Chat assistant	Proactive agent
Trigger	Only when a human asks	Continuous monitoring
Context	Single conversation	Must retain project state and history
Turns	Usually 1–3	A single task can take 10–30 turns
Cost curve	Linear and predictable	Compounds — every turn re-reads the whole history

In our agent cost analysis, a task adding 3,000 tokens of context per turn reaches 20x the first turn's input by turn 20. If an agent watches several repositories, multiply again.

Practical advice: run in off or confirm mode for a week, measure how often it triggers and how many turns each task takes, then decide whether auto mode is affordable.

Getting started

Find dsh-superagent on GitHub and read the accept-mode section
Trial it on a repository that does not matter, not on production
Write down explicitly when it may act and when it must ask — that is the prerequisite for auto
Set a spend cap to avoid runaway loops