What happens when the worker can delegate?

Termic 0.24 · July 24

The worker got a staffing desk

Termic 0.24 adds an experimental command line for creating, listing, watching, opening, waiting on, and archiving tasks in the desktop app. It is designed to be scripted: JSON and streaming JSON report state, heartbeats separate a busy agent from a stuck one, and exit codes are pinned.

The consequential detail is who can use it. Every task receives the command path and its own identifier. An agent can open more tasks, wait for them, and read what they produced. The release notes call out the new operating pattern plainly: one agent can farm a job to several agents in parallel.

The unit of work is no longer one conversation. It is a roster.

For interface designers, that changes the supervision problem. A chat transcript cannot carry the full state of a team. The product now needs clear role names, worktree locations, start and finish signals, and a place where parallel results return for comparison. Delegation makes information architecture visible.

AICreate Agent · July 22

Let the interface show why this tool was picked

AICreate's new agent sits above a catalog of 80 tools. On-device AI chooses the relevant local or connected Picsart tool, then runs it after approval. The update also adds a one-click arrangement for its Flow canvas and batch processing across local models.

A tool router is useful only when its choice remains inspectable. The agent can make the first match, but the surface should still name the selected capability, show the payload that will run, and give the reader a meaningful approval point. Think candidate recommendation, not invisible reassignment.

Agent-ready websites · July 13

A clearer site gets the agent hired faster

A new framework for agent-ready ecommerce sites separates readiness into machine readability, actionability, and decision reliability. Across 300 runs, the structured version reached 89.3 percent strict task success against 49.3 percent for the baseline. Average steps fell from 9.31 to 6.49.

That is a design result, not merely a model result. Semantic structure, explicit actions, and dependable product information reduce the amount of inference a browser agent has to perform. The best agent interface can still be an excellent human interface: fewer ambiguous labels, clearer state, and a path that does not require guessing.