ASSAY Try one repo

Conference overviews · Apsara Conference 2026 4 sections · partial review

assay · overviews · 2026-09-22–24 · Hangzhou

Apsara Conference 2026

Apsara ran in Hangzhou on 22–24 September 2026. This is a partial review: one keynote section based on published text, plus two interviews and one product talk reviewed from automatic English captions. It does not cover the full keynote or parallel-forum programme. The three captioned sections add evidence to the concept register; this page records what that limited corpus supports.

Reviewed 2026-10-02 · Official event overview · official playlist

§1The event in one paragraph

The reviewed material spans Alibaba’s infrastructure strategy, two creative-tool interviews and one model-routing product talk. The useful additions for agent engineering are narrower than the programme: a workflow can become a shared skill, visual references can structure generation, human taste remains necessary when producing an image or video becomes easier, and a model can be chosen at each call inside an agent task rather than once per prompt. These are vendor and customer accounts, with no controlled reliability, quality or cost evaluation.

§2The 4 learnings that matter

  1. A workflow can become a reusable skill

    WonderClip describes converting canvas workflows into team-shared skills exposed through an API. A published versioned format and cross-harness portability were not demonstrated. See reusable skills.

  2. Creative controls sit around the model

    Editable storyboards and 3D-stage references make intent more explicit before generation. They are conditioning inputs, not hard enforcement boundaries. See harness engineering.

  3. Generation does not supply judgment

    Magnific’s Jose Florido describes taste as something built through experience. The interview gives no measured cost or quality comparison and does not validate a particular review process. See human intent and judgment.

  4. Routing can happen at every call

    Alibaba Cloud International’s Jeffrey Zhang presents Smart Fusion as one model endpoint that chooses a model at each LLM call inside a task, because most sub-steps of even a hard task are simple. A small, fast decision model that does not generate long outputs makes the choice; the few steps the speaker calls critical go to several frontier models whose results are combined. The benchmark is self-run, the customers are unnamed, and prompt-cache effects are not discussed. See model tiering and routing.

§3Where Assay stands

Assay’s briefs separate stated intent from execution, and its role skills package reusable procedures. Those mechanisms are relevant to the interviews. Assay does not claim a visual workflow editor or video-generation capability, and this scan does not establish that creative conditioning provides an enforcement boundary. Assay assigns model tiers per desk role by convention, a coarser grain than per-call routing. A routing endpoint that hides which model answered could not serve a role whose verdicts must stay above a capability floor.

§4What this event added to the register

The three captioned sections contribute to four existing concepts: reusable skills, harness engineering, human intent and judgment, and model tiering and routing. No new concept is minted. The keynote’s model and infrastructure announcements remain background material rather than additional concept sightings.

§5Method and coverage

Four reviewed sections: one CEO keynote from Alibaba’s published English keynote text, two complete interviews, WonderClip (18:51) and Magnific (10:08), and one complete product talk, Smart Fusion (11:37), from automatic English captions. WonderClip was published on 2026-09-25, Magnific on 2026-09-24 and Smart Fusion on 2026-09-30; their exact recording dates are unconfirmed. The Smart Fusion talk was found when the playlist was re-listed on 2026-10-02; the other additions since intake were one promotional clip and short clips, excluded as promos. Its customer-case figures were read from the on-screen slide; the benchmark slide was not read, so no benchmark figure is repeated here. The conference’s remaining keynotes and parallel forums are not covered. Initial promotional and unavailable-video counts were an intake snapshot, not a current playlist census. Keynote roadmap and performance claims requiring further primary confirmation are not repeated here.

Every conference sighting in the register links the recording used for the review.