Zen Agents (by Zencoder)

Zen Agents are Zencoder's customizable, sharable AI agents that run on selected code or an entire file directly inside the IDE. You highlight a block, type a short command like /review or /unittests, and the agent executes its specialized task on that selection. Each agent is a saved configuration — a name, a command/alias, prompt instructions, and a hand-picked tool list — so recurring work like security review, release-note generation, API docs, or onboarding walkthroughs becomes a maintained command instead of a doc nobody opens. Configuration is deliberately narrow. Building an agent means filling four required fields: name, command/alias, instructions, and a per-agent tool allowlist. That allowlist is where the control lives — git, File Search, Full Text Search, Web Search, Fetch Webpage Content, Semantic Code Search, File Editor, Execute Shell Command, and Gather Requirements. Instructions steer behavior; the tool list caps blast radius, so a read-only reviewer never gets File Editor or shell access. Zencoder's own guidance is blunt about this: keep instructions focused and limit tool access. Admins can leave an agent Personal or flip it to Shared so the whole organization sees it, and All/Zencoder/Custom tabs (with counts) separate prebuilt vendor agents from your own builds. Custom agents can be submitted to the community marketplace through the zenagents-library GitHub repo. The docs also draw a clean line between agents and Skills: agents are interactive — you pick one and chat — while Skills are passive instruction packages the agent loads on its own from task context. The constraint is scope. These agents execute inside Zencoder's IDE surface, so the payoff tracks how much your team actually works there. If you're already on Zencoder, you get codified standards that travel as commands; if you're standardized on Cursor rules or GitHub Copilot custom instructions, those cover similar ground without switching tooling.