Devika

The good idea here was showing the agent's own state in a web interface, so you could watch it decide rather than reading a log afterwards. In early 2024 that felt like the future, and it taught a lot of people what these systems actually do between the prompt and the pull request.

It has been essentially still since 2025, and an autonomous agent that does not track model behaviour becomes a museum exhibit quickly. Pick OpenHands, which occupies the same ground with people still working on it, and keep this bookmarked for the history.

There is no gap between the claim and the reality to expose, because the claim was modest. What was promised was an experiment, what shipped was an experiment, and the experiment stopped. The risk for a reader is entirely secondhand: a name that circulated as the open answer to a commercial product carries expectations the project itself never made.

Anyone arriving from an article should read the repository's own description first. What it does right: labelling itself experimental before anyone else did, and never claiming otherwise while attention was at its peak.

The pipeline is worth stating because it was widely copied. 1. A high-level instruction is decomposed into an ordered list of steps. 2. Missing knowledge is fetched by browser automation against live pages rather than from a static index, which was a genuine advance in grounding. 3. Code is produced against that context.

The plan is a list, not a graph, so a step that invalidates an earlier assumption has no route back. Nothing checks the output. No benchmark was published, despite the project being positioned against a product that published one.

In 2024 a credible open clone of a heavily funded product looked like the beginning of a business. It almost never was. Attention arrived in weeks and dispersed just as fast, because there was no distribution to retain it, no data accumulating and nothing a user would find painful to leave.

No funding, no revenue attempt and no acquirer ever materialised. The pattern repeats often enough on this board to be a rule: viral parity with a funded product is a marketing event, not a moat. Position: none, and treat the wave itself as the lesson.

Deployment would mean a Python environment and a set of provider keys on every machine, with no central console, no access federation and no record of what any agent did on whose behalf. That is a compliance gap I cannot close with policy alone, and there is nobody to ask for the features that would close it.

It also cannot be scheduled, so it produces nothing our pipelines can measure. I have no objection to curiosity on a personal laptop with a personal key. As a supported tool, not yet.

The model layer was generous for its time: several hosted providers and a local option, so the whole thing ran on my own hardware with no account. That was not common in early 2024 and it is the reason I still have a copy.

MIT means the browsing loop is mine to take, and that loop is the interesting part, worth more as a reference than the rest of the codebase combined. There is no protocol support and no plugin system, so extending it means editing it. For a project this size that is fine, and it is the only option left.