feat(hack16): agentic search by starsirius · Pull Request #17544 · artsy/force

starsirius · 2026-08-05T14:41:11Z

The type of this PR is: Feat

Description

This PR includes the agentic search experience from the hackathon (with the commerce portion removed).

The artwork search results return mocked data. We'll fill in with the right search algorithm.
It is guarded behind the emerald_agentic-search feature flag, and if we want to, it's safe to merge and iterate.
The code was quickly put together for hackathon demo purposes, and can definitely use refactoring/polish.
There is no test.

Agentic loop considerations

It occupies a web process for the length of an agentic loop which could be multiple seconds, which could be real performance implications. Long-lived (and streaming) requests might deserve a dedicated service.
The endpoint will need to live elsewhere to be consumed by Eigen and other clients.
Streaming can be considered for better responsiveness.

Migration

-[x] Set the ANTHROPIC_API_KEY secret

erikdstock · 2026-08-05T14:45:56Z

+        {typeof artwork.priceUsd === "number" ? (
+          <Text variant="xs" color="mono100">
+            {formatPriceUsd(artwork.priceUsd)}
+          </Text>
+        ) : null}


this can't be handled by the MP Money type?

Eventually it should. This demo simply mocks the artwork search results and does not really query MP for data.

erikdstock · 2026-08-05T14:48:09Z

+      const response = await fetch("/api/advisor-agent/chat", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ messages: nextWire }),
+      })
+
+      if (!response.ok) {
+        throw new Error(`Request failed: ${response.status}`)
+      }
+
+      const payload = await response.json()


so the force backend handles this rather than mp - do we have a library like swr that can give us an easy loading state?

It seems we are interested in building this on Eigen. If that's the case, the agentic loop will probably be moved elsewhere to be shared.

erikdstock · 2026-08-05T14:50:42Z

+import { type NextFunction, Router } from "express"
+
+const ANTHROPIC_API_URL = "https://api.anthropic.com/v1/messages"
+const MODEL = "claude-opus-4-8"


curious if sonnet can do this just as well - maybe good to make this configurable via the ENV.

good idea! I probably won't add this to another env var, but it'd be helpful to evaluate.

…mmerce-poc

relativeci · 2026-08-12T12:30:16Z

#9574 Bundle Size — 9.03MiB (+0.06%).

df560be(current) vs b83648c main#9569(baseline)

Warning

Bundle contains 29 duplicate packages – View duplicate packages

Bundle metrics

2 changes

1 regression

	Current #9574	Baseline #9569
Initial JS	`2.94MiB`(`+0.01%`)	`2.94MiB`
Initial CSS	`0B`	`0B`
Cache Invalidation	`97.48%`	`97.48%`
Chunks	`111`	`111`
Assets	`114`	`114`
Modules	`5705`(`+0.02%`)	`5704`
Duplicate Modules	`691`	`691`
Duplicate Code	`5.05%`	`5.05%`
Packages	`248`	`248`
Duplicate Packages	`28`	`28`

Bundle size by type

1 change

1 regression

	Current #9574	Baseline #9569
JS	`8.8MiB` (`+0.06%`)	`8.8MiB`
Other	`232.77KiB`	`232.77KiB`

Bundle analysis report Branch starsirius/agentic-commerce-poc Project dashboard

^{Generated by RelativeCI Documentation Report issue}

- Keep a turn inside the app's 29s production timeout: cap the agentic loop at 4 steps and run the model at medium effort. - Require a signed-in user and bound the client-supplied transcript, so the endpoint isn't an open proxy to Artsy's Anthropic account. - Raise max_tokens to 16000 (thinking shares the budget) and pass the stop reason through so truncated, refused, and step-exhausted replies say so instead of rendering as finished answers. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The chat endpoint requires a signed-in user, but the UI only gated on the feature flag — logged-out visitors could open the panel and type, and only learn they needed an account once the request came back 403. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

claude · 2026-08-14T08:56:30Z

+    return
+  }
+
+  const messages: AgentMessage[] = [...incomingMessages]


The client-supplied messages array is forwarded verbatim to Anthropic with no per-user or per-session rate limiting. Any signed-in user (once the flag is on) can supply arbitrary role/content blocks — including fabricated system-style instructions — and get a general-purpose Claude Opus proxy billed to Artsy's account. The per-request size bounds cap one turn, but nothing caps requests per user. Since the PR notes this "spends against Artsy's Anthropic account," consider a simple per-user rate limit before this graduates from the flag. Not blocking given it's flag-gated and login-gated.

claude · 2026-08-14T08:56:51Z

Code Review

Summary

Adds a flag-gated (emerald_agentic-search) “Ask an art advisor” chat widget to the Collect route. A new Express endpoint runs a short agentic loop against Anthropic with a single mocked search_artworks tool backed by a hardcoded demo catalog. The author flags this as hackathon code: mocked data, no tests, staging links, safe to merge and iterate behind the flag.

Issues Found

Reviewed against the hackathon-behind-a-flag framing, so nothing here is blocking.

🟡 Important — no per-user rate limiting on a metered endpoint (advisorAgentServerRoutes.ts:158)
Client-supplied messages go to Anthropic verbatim with only per-request size caps. A signed-in user can send arbitrary role/content and effectively use Artsy's Anthropic key as a general LLM proxy, with no cap on request count. Flag- and login-gated, so fine for the demo, but worth a per-user limit before wider rollout. Left inline.

🟢 Suggestion — failed turn desyncs transcript from wire state (AdvisorAgent.tsx:283-293)
On a non-OK response the user's message stays in transcript and an error bubble is appended, but wireMessages is never updated with that user turn. The next send builds nextWire from wireMessages, so the earlier question is dropped from the model's context while still shown on screen. Likely intentional (retry-friendly), but the visible/actual history diverge.

🟢 Suggestion — substring title match produces false previews (AdvisorAgent.tsx:116-129)
findRecommendedArtworks attaches a preview when the normalized title is a substring of the normalized reply. Short titles like “Town” (→ town) or “4:04” (→ 404) will match unrelated words in prose. Minor for a demo; a word-boundary or exact-name check would tighten it.

🟢 Suggestion — demo catalog links/images point to staging (advisorArtworkCatalog.ts:26,36,…)
link values are https://staging.artsy.net/..., so “View on Artsy →” and thumbnails will break in production. Already called out in the file's ⚠️ comment; noting so it isn't forgotten when the mock is replaced.

Areas Reviewed

SSR safety: Clean. AdvisorAgent returns null before any browser-global access, and navigator.clipboard / window.setTimeout are only touched inside the handleCopy event handler. No module-scope or render-time globals. Server route lives under Server/server.ts.
Typographic quotes: Rendered copy uses curly quotes (AdvisorAgent.tsx:358,441); straight apostrophes only appear in comments, the system prompt, and tool descriptions (not user-facing UI).
Feature gating: Consistent — client gates on useFlag + isLoggedIn; server gates on server-side Unleash + req.user (403), and featureEnabled falls through to 404 when off.
Testing: None, acknowledged by the author. Given the flag, acceptable for now.

Questions for Author

The endpoint occupies a web process for the full agentic loop (up to 4 Anthropic calls) against the 29s request timeout. Under concurrent use, has worker saturation been considered, or is the flag expected to stay low-traffic until this moves to a dedicated service (as the description hints)?

leamotta · 2026-08-14T09:37:10Z

@rquartararo do you think we can merge this as is, to play around on staging?

rquartararo

Excited to test this out!

starsirius added 4 commits August 5, 2026 10:28

feat: spike on agentic commerce experience

d5d0763

feat: remove purchase capability

f190910

refactor: promote advisor agent to a root component

1e9c675

chore: guard it behind a feature flag

8d4635f

starsirius added the hackathon label Aug 5, 2026

starsirius self-assigned this Aug 5, 2026

starsirius requested a review from a team August 5, 2026 14:44

erikdstock reviewed Aug 5, 2026

View reviewed changes

starsirius changed the title ~~feat: agentic search~~ feat(hack16): agentic search Aug 5, 2026

starsirius marked this pull request as draft August 6, 2026 06:40

Merge remote-tracking branch 'origin/main' into starsirius/agentic-co…

dedca12

…mmerce-poc

Uh oh!

Conversation

starsirius commented Aug 5, 2026 • edited by leamotta Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Agentic loop considerations

Migration

Uh oh!

erikdstock Aug 5, 2026

Choose a reason for hiding this comment

Uh oh!

starsirius Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

erikdstock Aug 5, 2026

Choose a reason for hiding this comment

Uh oh!

starsirius Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

erikdstock Aug 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

starsirius Aug 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

relativeci Bot commented Aug 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

#9574 Bundle Size — 9.03MiB (+0.06%).

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

claude Bot Aug 14, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot commented Aug 14, 2026

Code Review

Summary

Issues Found

Areas Reviewed

Questions for Author

Uh oh!

leamotta commented Aug 14, 2026

Uh oh!

rquartararo left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

starsirius commented Aug 5, 2026 •

edited by leamotta

Loading

erikdstock Aug 5, 2026 •

edited

Loading

starsirius Aug 6, 2026 •

edited

Loading

relativeci Bot commented Aug 12, 2026 •

edited

Loading