How to Use GDevelop AI Agent - First Feature in One Evening 2026
If you are searching gdevelop ai agent, you are usually choosing between two vibes: paste prompts into ChatGPT until a browser demo appears, or open GDevelop and let an agent that knows your project build events, objects, and behaviors inside the editor.
GDevelop’s official June 30, 2026 post puts it bluntly: chatbots made the idea → playable box step easy; they did not remove publishing, touch controls, multiplayer, or long-session context collapse.Why Use GDevelop When You Can Just Use AI? The built-in Agent is the answer that stays inside the engine—visible events you can edit after it finishes.GDevelop AI wiki
This is a one-evening first-feature path—not a Steam save-path freeze week (that challenge lives here) and not a Kimi coding guide (Kimi beginner path).
Who this is for: beginners, creators prototyping on Chromebooks or thin laptops, and indies comparing Agent vs chatbot stacks.
Time: 90–120 minutes for one mechanic you can playtest.
Honest limit: AI credits are weekly and plan-gated; results vary; you still own playtests and store disclosure.
Why this matters now
- Official “engine not chatbot” framing is current. The June 30 GDevelop blog argues mobile export, free multiplayer, templates, and project-aware AI beat pure vibe coding for anything past a coffee-break demo.GDevelop blog
- Ask AI docs are concrete. Auto edit on/off, reasoning levels, restore points, and weekly usage resets are documented—not folklore.Wiki
- Search still needs an evening checklist. Marketing pages explain the Agent; SERP lacks a gated first-feature playbook with credit honesty and Steam pairing.
Open-source demand stays loud: 4ian/GDevelop sits around ~25k GitHub stars—the Agent ships inside that same editor ecosystem.
What the GDevelop AI Agent actually is
| Mode | What it does | Use when |
|---|---|---|
| Ask / chat | Answers questions about features using your project context | Learning, debugging concepts |
| Build for me (Agent) | Creates/edits scenes, objects, behaviors, variables, events, official extensions/assets | Shipping a mechanic tonight |
| Auto edit on | Applies changes as it works (default) | You already agreed the plan |
| Auto edit off | Proposes changes; you confirm each one | Brainstorming without wrecking the scene |
Unlike ChatGPT pasting GDScript-ish guesses into a void, GDevelop AI is specialized for the engine and can read your objects, behaviors, variables, and events.Wiki It knows built-in features and official extensions—not arbitrary community packs unless approved.
Prerequisites
- A free GDevelop account (browser editor is enough).
- A project with one scene and a player object (template is fine).
- 90–120 minutes without Discord tab-fighting.
- Willingness to ship one mechanic (patrol enemy, coin score, door unlock)—not an RPG system.
- Optional: cloud project if you want automatic restore points before each AI request.Wiki
Beginner path - first feature in one evening
Step 1 - Write a three-line feature card
Example (copy and edit):
Feature: Enemy patrols between markers A and B, chases player within 200 px, returns when out of range.
Do not change: Player controls, camera, existing UI.
Done when: I can die once and restart without broken events.
Vague prompts like “make an RPG” waste credits. Official guidance: be specific with values.Agent blog
Step 2 - Save or duplicate before Auto edit
- Cloud projects: GDevelop saves restore points before AI requests—know where Project saved restore lives.Wiki
- Local projects: File → save a dated copy. Automatic restore is not available the same way.
Step 3 - Plan with Auto edit off
- Click Ask AI (top right).
- Turn Auto edit off.
- Paste your feature card and ask for a numbered plan only.
- Negotiate until the plan names objects, behaviors, and events you recognize.
Planning with Auto edit off saves weekly AI usage.Wiki tips
Step 4 - Build with Auto edit on
- Start a new chat for the build (one chat per task).
- Turn Auto edit on.
- Say: “Build the agreed plan. Do not change player controls or camera.”
- Watch step statuses (Exploring / Editing / …). Click Stop if it drifts.
- Prefer Medium reasoning first; raise only if the Agent fails twice.Wiki
Step 5 - Playtest like a stranger
Run Preview and check:
- Happy path (patrol → chase → return)
- Edge path (player never enters range)
- Restart path (scene reload does not duplicate enemies)
If broken, do not start a new mega-prompt. Stop, restore if needed, rephrase one failure.
Step 6 - Read the events (the learning tax)
Open the event sheet the Agent wrote. Rename cryptic objects. Delete dead events. This is how beginners become developers—marketing demos skip this step.
Developer path - G1–G6 gates before you call it done
| Gate | Pass when |
|---|---|
| G1 Scope | Feature card has numeric ranges and a “do not change” list |
| G2 Plan | Auto-edit-off chat ends with an agreed numbered plan |
| G3 Build | Auto-edit-on chat finishes without silent half-builds (or you Stopped and restored) |
| G4 Playtest | Three paths above pass on a cold Preview |
| G5 Credits | You logged remaining weekly % and did not burn High/Maximum on a coin pickup |
| G6 Disclosure | If Agent-built art/audio/narrative ships to Steam, store truth matches Steam AI disclosure keep/hold |
Receipt mindset (gdevelop_ai_agent_first_feature_receipt_v1.json):
{
"project": "your-game",
"feature": "enemy_patrol_chase_200px",
"auto_edit_plan_chat": true,
"auto_edit_build_chat": true,
"reasoning_level": "medium",
"playtest_paths": ["happy", "edge", "restart"],
"weekly_ai_percent_after": 72,
"steam_disclosure_needed": false,
"reviewed_at": "2026-07-23"
}When you later promote a Steam demo, pair this with GDevelop save-path help—Agent features do not fix Storage roots.
Credit budget - company / creator diligence
Official wiki notes:
- Usage shows as a weekly percentage near the message box; resets Sunday 23:50 UTC.
- Higher reasoning levels cost more; Medium is the default balance.
- Extra usage via subscription tiers, purchased credits, creator feedback, or promo credit claims.Wiki — Cost of AI requests
Company note: treat weekly AI % like a cloud budget line. Do not run Maximum reasoning on UI tweaks during a jam week.
Privacy note (from GDevelop): the AI is not trained on your individual projects; projects are only accessible while you use the conversation; it has no access to profile/leaderboard analytics.Wiki — AI privacy
Common mistakes
- One chat for everything — context mush; start New chat per feature.
- Auto edit on while brainstorming — accidental scene rewrites.
- “Make my game fun” — no numbers, no acceptance test.
- Skipping restore discipline on local projects — no undo story.
- Assuming community extensions work — Agent prefers official extensions.Wiki
- Shipping without reading events — you cannot debug what you never inspected.
- Confusing this with Steam save freezes — different job (7-day challenge).
Agent vs ChatGPT / Kimi - keep/hold
| Stack | Keep when | Hold when |
|---|---|---|
| GDevelop AI Agent | You want editable events inside GDevelop + one-click export | You need custom C++/GDExtension pipelines |
| ChatGPT / Claude chat | Engine-agnostic brainstorming | You expect it to publish APK/iOS for you |
| Kimi / coding agents | You are writing engine code outside GDevelop | You wanted no-code event sheets |
For a prompt-first Android no-code overview, see How to Create a Complete Android Game with One AI Prompt. For coding-model weekends, use the Kimi guide.
Key takeaways
- GDevelop AI Agent builds inside your project—objects, behaviors, events—not a detached chat file.
- June 30 framing: chatbots speed prototypes; engines still own shipping and iteration.GDevelop blog
- Plan with Auto edit off; build with Auto edit on.
- One New chat per feature; click Stop when it drifts.
- Start Medium reasoning; raise only after two failures.
- Cloud restore points ≠ local projects—save copies yourself.
- Weekly AI % resets Sunday 23:50 UTC—budget it.
- Read and rename Agent events before you call the feature done.
- Steam labels are separate—use the disclosure keep/hold when player-facing AI content ships.
- Save-path Steam week is a different URL—do not merge jobs.
FAQ
Where is the GDevelop AI Agent?
Open your project and click Ask AI in the top-right corner. Use Build / agent-style requests when you want the project edited; use chat when you only want answers.Wiki
Does the Agent replace learning events?
No. It accelerates scaffolding. You still verify Preview and read the event sheet. That is the dual-audience bar.
How do I avoid burning all weekly credits?
Plan with Auto edit off, keep chats focused, stay on Medium reasoning for small features, and stop runaway builds early.Wiki
Can it use every Extension Store pack?
It knows built-in features and official reviewed extensions. Community extensions may not work unless approved.Wiki
Do I disclose Agent work on Steam?
If the Agent only wrote events you would have written by hand, treat it like tooling. If it generated player-consumed art/audio/narrative that ships, follow Steam Content Survey honesty—see Steam AI disclosure debate.
Is this the same as the GDevelop Steam save-path challenge?
No. That challenge freezes Storage paths before fest upload. This article ships your first Agent feature in one evening.
Conclusion
The GDevelop AI Agent is useful when you want a project-aware builder, not another chatbot that forgets your scene graph. Lock a three-line feature card, plan with Auto edit off, build with Auto edit on, playtest three paths, and read the events. Bookmark this evening path—then, when Steam week arrives, switch to the save-path freeze and keep official Ask AI docs open while you work.