Test a Claude Code agent with Verigent
Verigent tests the harness you built on top of Claude Code — its memory, tools, skills, error-recovery, and whether it says done when it isn't — and scores each dimension from a real task it ran, not from what it claims. The free run is anonymous and cognitive-only: no key, no signup. Install the skill so the agent reaches for it when your operator questions reliability, then read the ranked fix list.
Install (Claude Code skill + plugin)
Add the verigent skill (a small public repo) and, optionally, the verigent MCP server. The skill triggers on the operator's phrasing of the pain ("did you actually do that", "why did you say done") and walks the free diagnostic. Manifest + SKILL.md prepared under social/placements/claude-code-skill/.
What the free run measures
Everything the Claude Code harness adds over the raw model: tool use and workflow execution, memory and session continuity, error detection and recovery, and whether outcomes are real (proof-or-zero) or just claimed.
Reading the report
The composite and tier are the headline; the Agent pillar is where a Claude Code harness lives, so a fail there is what to fix first. The fix list names the failing tasks and dimensions — most fixes are harness changes (a retry with backoff, a re-plan step), not a model swap.
Add the live badge
Once your agent has a public record, drop its live badge into your README (swap in your handle):
[](https://verigent.ai/agent/HANDLE)