AI research wiki
agentic-systems · 2026-11-02 · Tier 2

DeepSeek Super-agent and the fresh-context reviewer pattern

DeepSeek Super-agent and the fresh-context reviewer pattern

Sources: DeepSeek Harness team, "Super-agent" technical post and open-source prototype (weekend of 2026-11-01), via The Information and X summaries; Google's open-sourced Elephant-Goldfish plugin; a widely shared summary of the Claude Code team's review setup; Claude Code role-split settings thread Links: DeepSeek prototype repo · The Information · Elephant-Goldfish repo · threads: @MinLiBuilds, @omarsar0, @VaibhavSisinty, @undefinedKi, @damkina7 Raw: raw/rss/2026-11-02-the-information-deepseek-unveils-autonomous-coding-agent-prototype.md, raw/twitter/feed/2026-11-0*-ranked.json

TL;DR

DeepSeek's argument: with one agent task the bottleneck is the model, but with many tasks the bottleneck is the human, who can follow only 3 to 5 sessions at once. Super-agent splits the system in two. A chat-only Router talks to the user and dispatches work; isolated Workers run in parallel in the background. Event-driven triggers, full-duplex interruption and an external watchdog handle scheduling and recovery. Reported p95 task completion fell from 29.5 to 15.7 minutes. DeepSeek calls it a research prototype; several ideas in the post are future directions. On the same weekend, two other designs made the same split between a long-memory coordinator and fresh-context workers: Google's Elephant-Goldfish (one long session drafts a spec; fresh subagents that never saw the conversation explain it back and attack it) and the Claude Code team's review setup (a second agent reviews the diff in a fresh context; parallel bug hunters plus a voting group; reported merged PRs more than doubled and reverts fell over 40%, second-hand numbers).

The coordinator keeps the memory, the workers keep the context clean
Three designs, one shape: isolation is the error-catching mechanism.
flowchart LR
  U["User<br/><small>3-5 sessions max</small>"] --> R["Router<br/><small>chat-only, dispatches</small>"]
  R -->|task| W["Workers<br/><small>isolated, parallel</small>"]
  W -->|result| V["Fresh reviewer<br/><small>no shared context</small>"]
  V -->|reject| W
  V -->|pass| O["Done<br/><small>p95 29.5 to 15.7 min</small>"]
  D["Watchdog<br/><small>external recovery</small>"] -.-> W
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class U input
  class R,V loop
  class W core
  class O exit
  class D err
  linkStyle 3 stroke:#f08c00,stroke-width:2px
  linkStyle 4 stroke:#2f9e44,stroke-width:2px
Blue is the human, amber is routing and review, purple is the workers, green is the exit, red is the recovery path.

Key points

  • Super-agent (DeepSeek). Router + Worker two-layer design; event-driven, interruptible, watchdog-supervised. Original long post is in Chinese; the open-source prototype is on GitHub.
  • Elephant-Goldfish (Google open source). For Claude Code, Codex and Antigravity. The Elephant (long session) writes a spec under a no-code rule. Goldfish subagents check it: one must explain it back (a wrong explanation means an ambiguous doc), others attack it, a last one checks that paths, structures and signatures are pinned. README says skip it for small, clear fixes.
  • Claude Code review (per summary). Tests, build and linter output shown before "done"; reviewer in fresh context; REVIEW.md per code owner; findings posted only if enough parallel agents agree; every finding fixed or closed with a reason. Reported: substantive review comments rose from 16% to 54% of PRs.
  • Role-split settings (practitioner thread). Opus plans and unblocks, Sonnet writes, Haiku scouts file:line, an Opus critic reviews the diff. Model routing by role inside one harness.
  • Same weekend, the token side. HarnessMM (covered 10-27: the model writes future commands with pass checks so the harness skips LLM calls whose next step is already decided) resurfaced with DeepSeek Harness numbers: 87.93% fewer tokens and success 71.4% to 85.7% on 98 HumanEval tasks.

Relation to prior wiki pages

  • N-of-a-kind: three coordinator/fresh-context designs in one weekend (Super-agent, Elephant-Goldfish, Claude Code review). With JevAct (11-01), which puts a separate gate in front of tool calls, the pattern is that error-catching now lives in a separately-contexted component, not a bigger model.
  • Extends DeepSeek Harness v0.1 (08-14), the append-only, plugin-everything harness DeepSeek open-sourced while repricing cache hits 6x. Super-agent is the multi-session layer on top.
  • Routing frame: Router-plus-cheap-workers is the model-routing shape the wiki tracks, moved up to the session layer. Fresh-context reviewers also cap cost because they run short contexts. See agent harness engineering and multi-agent systems.

Gaps

  • No public ablation separating the gain from parallelism versus from the router.
  • The Claude Code numbers come from a summary thread without a linked original.