DeepSeek dsh superagent Router Worker Architecture Forces Parallelism Cuts p95 47 Percent

DeepSeek Harness team published dsh superagent on October 31 2026 as an MIT licensed TypeScript bundle for DeepSeek Harness It is a research prototype not a product The README leads with this warning The design essay Exploring Super-agent Design Towards a More Autonomous Agent System is Chinese only the English summaries are AI generated

The architecture splits a coding agent into two layers a long running Router that handles all human interaction and dispatch and isolated Worker sessions that execute tasks in parallel Each Worker gets its own session directory and tool set The Router never executes work It cannot even run a shell command That structural constraint is the entire point

In DeepSeeks synthetic scheduling experiment the Super agent completed work requests at p95 15.7 minutes versus 29.5 minutes for a single session agent with a subagent tool The baseline never delegated once in five runs under a neutral prompt The speedup is not model capability it is architecture forcing parallel dispatch regardless of what any single model session decides to do

The bill for that determinism shows up in tokens Super agent reads 23.2 million cached tokens per run versus 6.1 million for the neutral single session baseline Nearly 4x the cache reads Removing the subscription pass through mechanism raises Router decisions 30 percent and cache reads 37 percent to 31.8 million without changing completion time

Source DeepSeek Harness team design essay section 7.1 Five runs per configuration Synthetic coding work simulated with a time consuming tool Measures scheduling and coordination not code quality

A TypeScript bundle v0.1.0 MIT license one commit on October 31 2026 Repository deepseek ai dsh superagent prototype Runs as a dedicated DeepSeek Harness profile called superagent Listens to GitHub notifications for one repository Optional Feishu or Lark bot for chat requests Returns results to the original conversation usually as a pull request Local web panel shows only items needing human attention

Not a standalone agent Not a model Not a cloud service Not production ready The deployment guide is explicit both presets run with danger full access as the operators OS user A worker can execute programs by absolute path and read the operators home directory The safety model is a dedicated Unix account not the prompt and not the shims

The Router runs the dsh superagent router preset Deliberately small tool set No shell No job wait caps at a few seconds It polls GitHub evaluates new triggers answers follow ups while workers are busy

Workers run the dsh superagent worker preset Full tool set One session one directory per Ticket Foreground shell commands background after five seconds so a human can still interrupt State lives in a single SQLite database under a stateDir you choose An Outbox routes replies back through the original channel

Six steps all recorded in the store

1. Source calls receive trigger Intake stores it once answers stored duplicate or filtered
2. Trigger goes to tickets already subscribed to its thread or queues an Eval on its primary thread
3. Assessment effect hands the candidate to the Router session which calls decide
4. Accept creates the Ticket its Mandate an Assignment the thread subscriptions and an acknowledgement in one transaction
5. Launch effect starts the worker session worker id in its own directory delivery effects pass it further input
6. Worker reports return through relay effects replies go out through the Outbox and the source CLI

A ticket is done when none of its pull requests remain open A pull request belongs to a ticket when the last path segment of its branch matches the tickets short name The recommended prepare command names branches agent DSH SUPERAGENT NAME

After a ticket starts later messages on the same thread are delivered directly to the subscribed worker The Router is not re reading every pull request comment Removing pass through changed completion time barely at all but cost roughly 30 percent more Router decisions and about 37 percent more cache read tokens 23.23 million to 31.77 million per experiment

This is the only mechanism in the prototype that reduces token cost The Router alone stress test found pushing trigger arrivals from about 1 per minute to 40 per minute only moved per request latency from 2.5 to 3.9 seconds because queued triggers batch into one call

Three accept modes per source per sender per message combination

Policy tables mark every combination as explicit inferred or ignore plus an accept mode Only allowed people can start work explicitly Everything else is at most a candidate By default candidates are only observed Work nobody asked for starts in observe mode that only records what the agent would have accepted

Synthetic stream of about 300 group chat and CI messages per run Roughly 20 real work requests Five runs per configuration Coding replaced by a tool that only consumes time Measures scheduling and coordination not code quality Message intervals and work durations scaled to one quarter Model call times not scaled Reported minutes should not be treated as production estimates

With a neutral prompt use a subagent for a piece of work when appropriate the model never delegated once in five runs It executed everything in the main session absorbed interruptions from new messages the whole time and posted that 29.5 minute p95

With a prompt rewritten to push delegation hard it delegated in four of five runs and got faster but judged the work too small to delegate in the fifth and produced the long tail 29.0 minute p95

Whether a multi agent setup works at all depended on prompt wording and in the moment model judgment The dual layer removes that coin flip structurally The Router does not execute Each Worker only sees the thread it subscribed to

Keeping many isolated sessions alive reads more cached context than one session The ablation row quantifies what the pass through mechanism buys Removing it changed completion time barely at all but cost roughly 30 percent more Router decisions and about 37 percent more cache read tokens

No independent reproduction of the experiment has been reported Three independent summaries offloop CellCog aibase report the same figures All vendor reported

Separate stress test of the Router alone trigger arrivals from about 1 per minute to 40 per minute Average batch size grew from 1.0 to 2.8 Amortised inference time per decision fell from 6.4 to 1.6 seconds Processing several events together made better use of model calls

dsh superagent runs on DeepSeek Harness which runs on Cordis Cordis is a plugin framework developed since 2022 as the kernel of Koishi a cross platform chatbot framework Its author Shigma Yifan Shi co wrote A Programming Paradigm for Spatiotemporal Composability with DeepSeek researchers posted to arXiv August 26 2026

Temporal composability revertible effects Everything a plugin does on load registers a command hooks an event opens a connection is an effect on the shared context Every effect carries its own inverse which the runtime tracks When the plugin unloads the system runs those inverses and reverts its effects No orphan listener no open connection no ghost command left behind

Spatial composability reactive coeffects A coeffect is what your plugin needs from the system a database a model a service In Cordis those needs are declared The runtime uses the declaration to manage lifecycle It does not activate the plugin until what was declared is served It reacts when dependencies change deactivating or reconfiguring the component when appropriate

In a traditional agent framework if an agent writes a new tool during a long running task it cannot mount that tool without restarting its process Restarting wipes ephemeral memory drops open network sessions resets execution state

Under Cordis the agent writes a new TypeScript tool plugin The harness loads the plugin into the live context at runtime Cordis checks the plugins coeffects mounts the tool and automatically updates the system prompt assembly for the next step If the tool fails or finishes its purpose the agent unloads the plugin Cordis unwinds the effect tree with zero leftover memory or socket leaks The agent modifies its own runtime in flight while preserving active session history

emit fire and forget waterfall short circuit on first truthy parallel all at once serial one after another bail short circuit on first non null Event names are registered through declaration merging This is how Router Worker communication works without tight coupling

Both presets run with danger full access as the operators OS user Commands run without approval The deployment guide says outright the wrappers establish defaults not a security boundary A worker can run a program by absolute path and read the operators home directory

Authenticates with a documented static header x orchestrator request 1 plus a localhost check Any local process that knows the header value gets panel permissions The essay concedes that if the cloud loop is compromised every machine connected to it is at risk

Dedicated Unix account containing only the agents credentials Separate machine or container Blocked commands configuration Pin the Harness version record the prototype commit test upgrades on a disposable environment keep a recovery path for unfinished Tickets A database migrated by a newer version cannot be opened by an older one Rollback requires the pre upgrade backup

Key technical specifications and benchmarks below.