Loading ToolBench
OS for your agents - a platform for building, deploying, and managing AI agents with MCP server integration
AgenticOS exposes 2 browser automation tools with detailed, LLM-optimized descriptions (150 - 200 chars each). Both tools have clear input schemas with typed parameters and descriptions. However, output schemas are not documented in the visible code, and neither tool declares error handling guidance, recovery paths, or idempotency guarantees. Tool names are action-verb-based (browse_*), which is good. Parameter descriptions are concrete and actionable. The main gaps are missing output schema documentation and no error classification or recovery guidance, typical of agent tools that delegate to external services without surfacing failure modes.
Work through a web page towards a goal, one chosen action at a time. Opens start_url in a real browser and repeats: read what is on the page, choose one of the things actually on it, do that. It chooses from the page's own elements rather than composing an action, so it is steady on forms, filters, consent gates and multi-step flows, and it stops and says so when a page offers nothing that reaches the goal.
Delegate an open-ended web task to an autonomous browser agent. Give one self-contained goal in natural language; the agent drives a real browser on its own - navigating, reading, clicking and extracting - and returns a text result. Prefer it when the page layout is unknown or the task needs judgement, not for a scripted flow a direct request would do.
Output schemas not documented. LLMs cannot infer what fields browse_page and browse_web return, forcing them to guess downstream tool compatibility and data extraction logic.
No error handling guidance. Tools delegate to external browser agents but do not document failure modes (timeout, page not found, goal unreachable, auth failure) or recovery paths. LLMs cannot self-correct on failure.
No idempotency or side-effect declaration. browse_page and browse_web may perform irreversible actions (form submission, purchase, account changes) but do not warn agents or require confirmation. Agents cannot safely retry.
Parameter 'private' in browse_page lacks format/length constraints. No guidance on what 'sensitive values' means or how many/which fields it accepts. LLMs may pass entire credential objects or oversized payloads.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2026-07-28+ | v2 |
No pagination or result-limit guidance. browse_web returns 'a text result' but does not specify max length, truncation behavior, or how to handle multi-page results. Large responses may exhaust context.