Capabilities
Capabilities are AdaL's built-in skillsets. Each one bundles a set of tools with the instructions for using them well: browser use, deep research, video generation, the autonomy auditor, and human-sounding writing. You never turn them on or off yourself. The agent reads your task and loads the right capability on its own, mid-session.
Before capabilities, these features were agent modes: you ran /agent browser-use to switch into a mode before starting, and switching back meant losing your place. Modes are gone. Ask for what you want ("open a browser", "research this", "make a video") and the capability loads when it's needed.
The capabilities in AdaL
| Capability | What it does | Try saying |
|---|---|---|
| Browser Use | Drives a real browser: opens pages, reads what rendered, fills forms, clicks through flows, takes screenshots. Your Chrome logins can carry over. | "Open a browser and check our staging deploy" |
| Deep Research | Multi-source, evidence-led research with durable source documents and line-level citations. | "Research how our competitors price usage" |
| Video (preview) | Creates production-ready videos: generated clips, voice-over, subtitles, trimming, composing scenes. | "Create a 30-second product demo video" |
| Autonomy Auditor | Audits your codebase for what agents need to work autonomously: telemetry, tests, enforcement, agent setup. | "Audit our codebase for autonomy" |
| Write Like Human | Rewrites AI-sounding text so it reads like you wrote it, without changing what it says. Loads on its own for text a person will read: PR descriptions, commit messages, docs, comments, posts. | "Rewrite this README so it doesn't sound like AI" |
When a task calls for a capability, the agent loads it on its own:
Check what's available and active in your session with /capabilities.
How it works
Capabilities run on a skill-over-tools design:
- Nothing loads until it's needed. A capability's tools and instructions stay out of your session entirely until the agent asks for them. Only a short description (~40 tokens) is always present, which is how the agent knows the capability exists.
- The agent decides, not you. When your task matches a capability's description, the agent loads it with one tool call. The full instructions and tools arrive, do their work, and are cached from then on.
- Capabilities leave when their work is done. After compaction, a capability you no longer need is discarded, and one you are still using is loaded again. A session that finished browsing never pays for browser tools again. Long sessions stay light and fast.
- They combine. Several capabilities can be active in one session. Research a topic with deep research, then turn the findings into a video, without leaving the conversation.
- Provider-agnostic. The same mechanism works across Anthropic, OpenAI, and Google models, with no dependence on any provider's native tool-deferral feature.
Skills vs. capabilities
Skills are reusable expertise you or the community write and install: markdown files that teach the agent a domain workflow. Capabilities are AdaL's built-in skills, available on first run with no install. The pattern is the same one, applied to AdaL's own power features.