Taxonomy — 28 capabilities · 112 systems
Agent capabilities
Capabilities are atomic tags that describe what an agent can actually do — run code, edit files, drive a browser, take phone calls — independent of which product category it belongs to. A CLI agent and a cloud agent can share 'long-horizon tasks'; a voice agent and a digital employee can share 'human handoff'. The tags let you compare across categories rather than only inside them.
Tags verified against the catalog seed of 2026-09-25. Filter the agent index by capability to compare across categories.
Agent capabilities — What the agent itself does
Background runs /background-tasks
Asynchronous cloud or background execution you can walk away from.
10 agents →Browser operation /browser-ops
Drives a real web browser: navigate, click, fill, extract.
12 agents →Cited answers /citations
Returns answers grounded in identifiable sources.
08 agents →Code execution /code-exec
Runs code in a real environment or sandbox.
16 agents →Code review /code-review
Reviews diffs and pull requests with feedback.
04 agents →Desktop control /desktop-use
Pixel-level computer use — screenshots, mouse, keyboard.
03 agents →File editing /file-edits
Reads and writes project files, including multi-file edits.
35 agents →Git & PRs /git-pr
Branches, commits, and pull-request workflows.
09 agents →Human handoff /human-handoff
Escalates, transfers, or requests approval from a human.
28 agents →Incident response /incident-response
Alert, on-call, and incident-triage loops.
07 agents →Long-horizon tasks /long-horizon
Unattended multi-step work running minutes to hours.
26 agents →Model routing /model-routing
Selects or routes across multiple underlying LLMs.
13 agents →Persistent memory /memory
Remembers context across sessions or runs.
12 agents →Repo context /repo-context
Indexes or reasons over a whole repository, not just open files.
34 agents →Shell access /shell
Executes terminal/shell commands.
25 agents →Voice I/O /voice
Realtime voice or telephony input/output.
09 agents →Web research /web-research
Searches, browses, and synthesizes information from the web.
17 agents →Workflow builder /workflow-builder
Visual or no-code composition of agent workflows.
06 agents →Infrastructure capabilities — How it composes into a stack
Agent identity & auth /identity
Identity, credentials, or authorization for agents.
07 agents →Agent payments /payments
Payment rails, wallets, or charging for agent actors.
05 agents →Agent-to-agent protocol /a2a
Speaks or implements an agent interoperability protocol.
08 agents →API-native /api-native
Designed to be embedded and driven via API or SDK.
70 agents →MCP tools /mcp
Model Context Protocol client or server support.
24 agents →Messaging surface /messaging-infra
Email, SMS, or chat channels agents can send and receive on.
10 agents →Multi-agent orchestration /multi-agent
Spawns, coordinates, or composes multiple agents.
12 agents →Sandbox runtime /sandbox-runtime
Provides isolated execution environments as a service.
02 agents →Self-hostable /self-hostable
Can be deployed on your own infrastructure.
16 agents →Tool calling /tool-calling
Function/tool use as a first-class primitive.
12 agents →Frequently asked
What are AI agent capabilities?
Capabilities are atomic tags that describe what an agent can actually do — run code, edit files, drive a browser, take phone calls — independent of which product category it belongs to. A CLI agent and a cloud agent can share 'long-horizon tasks'; a voice agent and a digital employee can share 'human handoff'. The tags let you compare across categories rather than only inside them.
How is this different from the category taxonomy?
Categories answer 'what kind of product is it' — an AI IDE, a cloud agent, a voice agent. Capabilities answer 'what can it do' — execute code, call MCP tools, route between models. Categories partition the catalog; capabilities cut across it.
What do the two facets mean?
Agent capabilities are things the agent itself does — file editing, web research, voice I/O. Infrastructure capabilities describe how the system composes into a stack — API-native, self-hostable, agent payments, sandboxed runtimes. Most infrastructure entries are building blocks rather than end-user agents.
How are capability tags assigned?
Each system is tagged only with capabilities verifiable from its documented feature set in the catalog seed. Tags are not self-reported by vendors and are re-checked when the catalog is regenerated.