The Google AI Studio developer guide provides several task examples explicitly tied to Gemini 4 Argon: generating a single-file Web Audio drum machine, auditing a website with a Managed Agent, establishing an accessibility baseline for a web app and iteratively fixing it, using multimodal video to locate user sticking points, and having Argon orchestrate sub-agents. The original prompts and steps are preserved below; the guide examples are not presented as measurements by this project.
Suitable tasks: Complex front-end delivery, website performance and SEO audits, verifiable code fixes, multimodal user-testing analysis, long-running task planning, sub-agent orchestration, and turning one-off debugging procedures into skills.
Unsuitable tasks: Inferring success rate, cost, latency, or actual quality for arbitrary projects from these official examples; the guide provides no unified measured data for these tasks.
Model version: Gemini 4 Argon, model ID gemini-4-argon.
Client, Agent, or API: Gemini API Interactions API, Google AI Studio, Managed Agents, and Google Antigravity; the entry point differs by example.
Recommended reasoning tier and parameters: Start by clarifying the goal, milestones, and acceptance signals. In the orchestration example, the guide recommends thinking_level="medium" or "high" for Argon; task-specific parameters still require measurement.
The following prompts are from the guide and have not been rewritten or combined.
Build a single-file HTML drum machine and synth using the Web Audio API with a 16-step sequencer, tempo slider, and a live waveform visualizer.The page sends this prompt to gemini-4-argon with thinking_level: "medium" and prints interaction.output_text. The guide describes the expected deliverable as a runnable single-file HTML application, but does not provide the actual output code or acceptance result from that run.
We want to fix accessibility and layout issues across our web app, but we don't have automated checks yet. First, write a headless Chrome script that crawls every route in src/routes/, inspects the live DOM and accessibility tree, and logs all violations to record our baseline score. Group the failing checks into shared root causes across our shared components, then fix them one cluster at a time. Re-run your script after each change to verify that each single fix resolves the entire cluster with zero regressions. Keep going until all routes pass with zero violations.The acceptance signals for this prompt are: obtain a baseline first, group failures by shared root cause, fix one cluster at a time, rerun the checks after every change, and finish with zero violations and zero regressions on every route. The guide does not show actual scan results.
Investigate our current session to see where things stalled or required extra steps, and analyze why. Propose updates to our AGENTS.md or custom skills so we can improve our flow and take a more direct path next time.This is the guide's reflection prompt for turning stalls and extra steps in the current session into updates to AGENTS.md or reusable skills; the page provides no specific diff.
The guide publishes a complete Interactions API task input:
Audit https://web.dev for performance, Core Web Vitals, and SEO. Query Google's PageSpeed Insights API for both Mobile and Desktop strategies. Check search indexing with Google Search for site:web.dev. Format the output as a side-by-side scorecard table with prioritized fixes.This example configures agent="antigravity-preview-09-2026", agent_config.type="antigravity", agent_config.model="gemini-4-argon", and environment="remote". The expected output is a side-by-side mobile/desktop scorecard and a prioritized list of fixes. The page provides no actual scores, API responses, or repair results.
The guide recommends first asking Argon to raise clarifying questions about edge cases, success criteria, and key milestones; using the planning phase to test the architecture and approach; confirming that tools, documentation, and intent are available before execution; and then letting the model write code, run checks, iterate across files, and pause for confirmation before irreversible steps. This is workflow advice, not another complete prompt.
The guide recommends connecting a test suite, runtime logs, browser screenshots, or custom validation scripts; recording a baseline first; and repeatedly testing hypotheses until explicit pass/fail signals turn green. When no metrics exist, first have Argon analyze the project and establish a baseline. This corresponds to the accessibility prompt above, but the page reports no execution results.
The guide says Argon can accept images, audio, video, code, and text; dynamically jump to video timestamps; inspect frames at high resolution; and connect visuals with narration or console audio. Its concrete task directions include recording a screen share with narration and live logs so Argon can locate timing problems and fix components, or supplying a 20-minute user-test recording to find where users get stuck and patch the underlying components. It does not publish complete prompts, video contents, or repair outputs for these tasks.
The guide recommends having Argon decompose problems, design interfaces, and review the final result, while delegating code exploration, test scaffolding, leaf-module migration, and high-frequency checks to Gemini 3.8 Flash sub-agents. For complex interactive applications, Argon can also orchestrate specialized models such as Nano Banana, Gemini 3.8 Flash TTS, or Lyria. The page gives a Mars Rover landing physics simulator as an example that Argon and Flash sub-agents can build together, but it does not publish that project's complete prompt, code, or comparison data.
The page was published on October 13, 2026, and credits Geneviève Huskens and Jack Wotherspoon.
The Web Audio example explicitly uses gemini-4-argon; the task input and thinking_level: "medium" are visible. The output is read only through interaction.output_text, and the page does not include the actual returned content.
The Managed Agent example explicitly provides antigravity-preview-09-2026, the Argon model, a remote environment, and the complete task input; the expected output is a side-by-side website-audit scorecard and prioritized repair list.
The accessibility example publishes a complete acceptance target from a headless Chrome baseline through cluster-by-cluster fixes, rerunning after every change, and ultimately zero violations, but provides no measurement report.
The guide presents “long autonomous runs,” “self-verification,” “multimodal video understanding,” and “sub-agent orchestration” as Argon development advice. These are Google’s official guide claims, not independent evaluations.
This article does not repeat the 1M shared window, Long Decode, Thinking-parameter, or sampling-parameter migration rules covered in existing entries; it retains only this guide's distinct task prompts, acceptance loops, and product examples.
“Verify,” “zero regressions,” and “until the task is complete” are goals in the prompts or recommendations in the guide; they do not mean the page has demonstrated that the model achieved those goals.
The page does not provide a unified API version, tool permissions, dataset, success rate, runtime, cost, sample size, or human-review protocol; actual limits may differ across entry points.
The Managed Agent website audit depends on PageSpeed Insights, Google Search, and a remote Antigravity environment. Without those tools, it should not be treated as an ordinary text call.
The multimodal video examples do not fully disclose input format, video length, audio handling, timestamp precision, or repair validation. Do not turn the conceptual task descriptions into a verified Agent Skills workflow.
For reproduction, fix the entry point, model ID, complete prompt, tool permissions, project files, test scripts, and acceptance criteria separately for each example. The Web Audio example can be tried directly in the AI Studio Playground; the accessibility example should save the baseline, each cluster diff, and every rerun result; the Managed Agent example additionally requires a remote Antigravity environment and its web-access/API tools. This article records the task designs in the official guide and does not claim to have run or validated any example.
Gemini 4 Argon