8 November 2026

AI and agents, every day. Straight to it.

AgentsRadar

Hcompany launches Holo4 models that combine screens, code and tool calls

Holo4 works across graphical interfaces and application programming interfaces (APIs); its execution software adds memory for extended workflows and desktop shell access.

A task can lose its thread when an application programming interface (API) call hands work to a desktop interface. Holo4, announced by Hcompany on September 28, 2026, is trained to work through graphical interfaces, code, MCP connections to tools and APIs. The model stays the same as the execution surface changes.

The surrounding software matters too. Hcompany describes rebuilding its agent loop with memory for extended workflows and shell access on the desktop machine. The reported behavior therefore belongs to a model operating within that setup; downloading its weights alone does not reproduce the full arrangement.

The architectural recommendation is to record each interface transition: the observed state, the action taken and the evidence that it worked. For an internal evaluation, choose a task that actually crosses those boundaries. A purely visual demonstration would leave the central coordination question unanswered.

Hcompany publishes benchmark trajectories that readers can inspect. Use those records to examine failure patterns before choosing a deployment. The operational decision remains: who verifies that a screen action produced the state subsequently reported by an API?

Source: Hcompany's Holo4 announcement, by Maxime Theillard and coauthors · Hcompany.

Rafael Moura

Agents & automation.

Read articles by Rafael MouraEditorial policy

Report an error

Use 10–1,000 characters; a reply email shares this space.

Next in your reading list