SI Agent News
DevelopingAgents

A prototype turns vibe coding into a 3D drive where AI options are forks in the road

A dev.to author proposes steering an AI coding agent through a 3D road map instead of a chat, and admits the first version faked its own progress.

86%share of one recorded run the author estimated was fake pacing

The short version

Why it matters

Chat boxes hide how an AI's options relate to each other, and that makes it hard to judge what an agent is doing. This prototype argues that an interface should show real alternatives and real evidence of work. Its own audit shows how easily a smooth animation can pass for progress.

A road map instead of a chat log

Today's vibe coding, where you describe software in plain language and an AI agent writes it, mostly happens in a chat. You ask for a cleaner layout, the agent writes code, and it says it is done. As dev.to reported in this post, the author's complaint is that a chat log struggles to show how directions relate: which are parallel alternatives, which depend on earlier steps, and which were tried and rolled back.

The proposal is to turn those relationships into space. The agent's candidate next steps show up as signposts and forks on a 3D road. You pick one, the agent executes it, and the result is pinned to the spot where you arrived.

How a question grows a road

The author's example is a request to build the home page of a weather app. Instead of one reply, the screen offers separate routes: design the information structure first, build a runnable page first, set the visual style first, or check the weather data source first. You can read the reasoning and steps for each, or type a direction of your own.

A component called a forecaster generates the candidates and estimates how different they are. Options that barely differ render as a nearly straight road, and options that differ meaningfully fork. After a step finishes, new directions open ahead. The prototype also has side-by-side comparison and keyboard route selection.

The idea is that you stop front-loading every decision into one long prompt. You see a few real choices at each step and steer as you go.

What is real and what is a demo

The post labels its visuals in three tiers: AI-made concept art, interface demos with locally injected candidates, and recordings of a live run against a real backend. That honesty matters, because many screenshots are demo data. The live run uses a small local model served through Ollama, a tool for running models on your own computer. Its text is genuinely streamed as the model produces it.

The direction forecasts in that run came from a fast local channel and are labeled "local preview" in the interface, so they are not a full model's forecast. The author also says this is not yet a full toolchain for editing code, running tests and verifying results. Those are design goals.

The audit that found fake progress

The most useful part of the post is a self-correction. A user told the author that nothing substantial seemed to be happening. The author then audited the older execution chain and found that steps labeled things like "retrieval" and "validation" did no real work. They updated progress after a built-in pause of 0.09 and then showed templated text. The only real model call came at the end.

The recorded events ticked at a suspiciously regular pace of 90–100 ms. For one specific recorded run, the author estimated that this fake pacing made up roughly 86% of the total time. The author stresses that this figure applies only to that diagnostic run.

Version v0.9.12 deleted the empty steps and kept a single real model request with streamed output, crediting only one item: the output. On re-testing, the regular timing pattern was gone.

Why a good-looking interface can mislead

The author's lesson is that the more convincing the visuals, the stricter the rules must be about what counts as progress. A road segment should stand for a chain of input, action, result and evidence. If no tests ran, arriving at the destination should not read as "verified."

The post adds that forks only help if the options truly differ. If two routes differ only in wording and execute the same way, the fork carries no decision value. To check this, the project groups near-synonyms such as "look up references" and "search for references" into one road. The author calls the mismatch between meaning and layout a hypothesis, not something tested with human participants.

What to expect next

Nothing here is a product you can install. It is a concept with a working prototype and a candid account of its flaws. The next step the author describes is connecting the roads to real development tools, so that a drive would show which files changed, which tests ran and where something failed.

Whether people prefer steering through space to typing in a chat is untested. The post makes the case, and it also gives a standard any agent interface could be held to: show only progress that actually happened.

What to watch

Try it

Next time an AI agent reports "done" or "validated," ask it to show the files it changed and the test output. Treat any progress label with no evidence behind it as decoration.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by dev.to · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news