What are AI agents — and how they work on Overclock
AI agents are instances of AI models that execute tasks autonomously: they receive a goal, plan the steps, use tools — terminal, files, browser — and deliver a result at the end, reporting what they did.
The difference between a chat and an agent is the difference between asking and delegating. Chat responds; an agent acts. This article explains what an AI agent does on its own, how several work together without turning into a mess — and how that runs, in practice, inside a cockpit.
What an AI agent does on its own
An AI chat returns text: you ask, it answers, and execution stays with you. An agent closes that loop — it acts. It receives a goal (“fix this bug,” “build this page”), decides the steps, uses real tools — edits files, runs terminal commands, tests what it did — and comes back with a result, not a suggestion.
The loop inside is always the same: goal → action → observation → next action, until the task closes. The agent carries the project's context along the way and reports what it did at the end. That's why the AI agent became the unit of work of vibe coding: you say the intention, it runs the loop.
On Overclock, an agent is a combination of four pieces:
CLI + model
The engine (Claude Code, Codex, Gemini…) and the brain running on it — each agent with its own, chosen per task.
Skills
Instructions that become capabilities: the design agent carries design skills, the QA one carries testing skills.
Effort
How much the model thinks before acting. Mechanical task, low effort; hard decision, high effort — the token in the right place.
How several agents work together
One agent solves one task. A delivery — a site with copy, design, and QA — has multiple fronts, and that's where it gets complicated: agents running loose in separate windows can't see each other, and the human becomes the messenger between them. The fix isn't a bigger agent — it's coordination.
Maestro delegates
A coordinator splits the request across the right agents — and never executes anything directly. Coordination and execution never mix.
Each in its own pane
Each agent works in a visible pane, with live status and cost. You see the whole team in a single grid.
Handoff
When an agent finishes, the maestro is woken up right away, reviews the delivery, and passes the baton — nobody sits in a waiting loop burning tokens.
The arsenal, the harness, and the recipes
On Overclock, all of this comes already built. The arsenal brings the Overclock Originals — ready-made agents, each with CLI, model, skills, and effort calibrated for its role. The harness is our doctrine in code: every task ships with the right bundle — right agent, right model, right effort, right skill — because the same model delivers more or less depending on how it's used. And recipes are ready-made harness formations: pick one and the wizard fills itself in.
Coordination runs on 52 native MCP tools — spawning panes, delegating tasks, reading results, syncing the team — without you installing anything. The full agent roster is at /features/agentes; if you still need the map of the category, read what is an ADE. And the fastest way to see agents working together is to pick a squad by the result you want and run it.