Jul, 134 min to read
Frontend
News · Google's Deep Research agent now emits HTML charts and streams its reasoning
Apr, 214 min to read
FrontendGoogle's Deep Research agent now emits HTML charts and streams its reasoning
The Gemini API's research agent splits into a fast interactive tier and an asynchronous Max tier, and for frontend teams the notable change is what the agent now returns: in-line charts, live thought summaries, and a reviewable plan.
Two agents, two latency budgets to design around
Google split its December preview into two agents that behave differently enough to force a frontend decision up front. Deep Research is positioned for "research experiences integrated directly into interactive user surfaces where lower latency is desired," and it replaces the preview release with what Google describes as reduced latency and cost at higher quality. Deep Research Max, by contrast, "leverages extended test-time compute to iteratively reason, search and refine," and Google's own example is a nightly cron job producing due-diligence reports by morning.
That distinction is a UI contract, not just a pricing tier. A synchronous surface where a user waits for a result should target the fast agent; anything Max-scale belongs behind an async job with a notification or a report inbox. Building the same waiting-state UI for both would misrepresent how long each is meant to run.
The agent now returns renderable charts, not just prose
The most concrete change for anyone building the receiving end is that Deep Research "no longer just creates text; it natively generates high-quality charts and infographics in-line with HTML or Nano Banana." Google frames these as "presentation-ready" visual elements dynamically produced from the underlying data.
HTML output means the frontend is now responsible for sanitizing and safely embedding model-generated markup, and for handling image outputs (Nano Banana) alongside text in the same report stream. This is a different rendering pipeline than displaying a markdown summary. Teams will need a layout that interleaves narrative text, generated charts, and citations — and a policy for what to do when the generated HTML is malformed or unexpected.
Streaming reasoning and a review step change the interaction loop
Two features reshape the moment-to-moment interaction. Real-time streaming exposes "the agent's intermediate reasoning steps with live thought summaries," delivering text and image outputs as they are generated. Collaborative planning lets users "review, guide and refine the research plan generated by the agent before it begins execution."
Together these define a two-phase interface: a plan-approval stage before the expensive work starts, then a live progress feed while it runs. For long-horizon research, that plan review is the leverage point — it lets a user redirect scope before compute is spent rather than discarding a finished report. Frontend work here is less about a chat box and more about an editable plan object plus a durable stream of intermediate states.
What this means for teams wiring research into a product surface
The agent's data reach now depends on configuration the frontend has to surface honestly. Deep Research can search "the web, arbitrary remote MCPs, file uploads and connected file stores — or any subset of them," including the option to "turn off web access entirely to exclusively search over your custom data." With MCP partners named as FactSet, S&P Global, and PitchBook, users in regulated fields will need to see and control which sources an investigation touched, and every report ships as "fully cited" output that the interface must render as verifiable links, not decoration.
The practical takeaway: this release moves the integration surface from "call an API, show a summary" to "manage a plan, stream intermediate state, render mixed HTML and image output, and expose source scope." It is available today in public preview on paid Gemini API tiers via the Interactions API, so the frontend patterns for these three phases — plan, stream, cited report — are what a team should prototype before committing to a single view.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.