SuperAgent

The DeepSeek Harness team has published a SuperAgent design article and GitHub repository, released under the MIT license and explicitly labeled as a research prototype.

SuperAgent's proactive action strategy: upon seeing suspected defect feedback, first collect logs and attempt to reproduce, and submit results only after confirming that action is needed; however, operations such as payments, external sending and modifying production data should not be executed on speculation alone.

In the SuperAgent prototype, a subscription pass-through mechanism lets subsequent events in the same discussion thread be delivered directly by the program to the associated Worker after a task is established, bypassing the global Router to reduce latency.

Subscription is only a routing relationship, not a permission; receiving a message does not mean the task scope can be expanded, and unauthorized operations must be rejected or authorization must be re-requested.

If the same discussion thread is mistakenly treated as the same authorization domain, message-driven privilege escalation occurs: appending sensitive operation messages to an existing discussion thread may induce a Worker to exceed its authority.

SuperAgent proposes Inline Spill and chained recall for long-term context management: the current context retains only summaries and object identifiers, while full content is stored in queryable locations and read level by level along associated objects when needed.

Enterprises deploying such agent systems should start pilots with read-only tasks (such as updating difference reports when new materials arrive), test with historical samples, and then connect to real events on a small scale.

The SuperAgent architecture adopts a Router-Worker two-layer design: the Router is responsible for global coordination and dispatching, while Workers execute specific tasks concurrently in independent contexts.

SuperAgent uses fast and slow paths for event filtering: the fast path performs deterministic checks (source, whitelist, etc.), the slow path has the model make semantic judgments, and the outputs fall into four categories: ignore, log, accept the task, and clarify.

Model judgment cannot replace identity authentication, access control and pre-execution permission checks; deterministic rules handle clear-cut issues first, and the model only handles candidate events that require semantic understanding.

The official deployment documentation states that Workers execute with operating-system user permissions, recommends isolating them in a dedicated account, and warns that the prototype must not be connected directly to a company's production repository.

In simulated tests, SuperAgent's median task completion time was 12.4 minutes and its 95th percentile was 15.7 minutes; under a neutral prompt, the 95th percentile for a single session plus a SubAgent was 29.5 minutes.

In the above benchmark tests, coding work used a stand-in tool that only consumes time, so the tests measure scheduling, concurrency and interruption management, not the actual speed, quality or success rate of the model in completing code.

The control Agent under the neutral prompt never proactively delegated subtasks in any of the 5 tests.

SuperAgent adopts a Router-Worker architecture that separates message triggering, task-acceptance decisions, long-term work items and actual execution: the Router decides whether to accept a task and whom to assign it to, and the Worker completes the task in an independent session and directory and writes back the result.