> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.rightbrain.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.rightbrain.ai/_mcp/server.

# Evals and Run Operations

> Replay file-backed evals, inspect agent run activity and costs, and use Azure OpenAI deployments.

## What's New

### File-Backed Task Evals

Task evals can replay reference runs that used direct image or PDF inputs. Stored processed values are reused, and direct model-facing files are restored from Project storage for the candidate run. See [Evals](/docs/production/evals).

### Detailed Agent Run Operations

Agent run views now combine total cost, status, failure reasons, phase timing, tool and skill activity, approval duration, actor identity, and the connection account used by each integration call. Agent lists also support server-side name and creator filters with grid and table layouts.

### Azure OpenAI and Claude Fable 5

Tasks can use supported Azure OpenAI deployments, and Claude Fable 5 is available for demanding reasoning and agentic workloads.

## Improvements

* Leaving an agent page no longer cancels its live run; reloads warn before disconnecting an active stream
* Deleting an integration preserves historical revision references and reports visible blocking dependencies
* Approval resumes preserve the original run actor identity
* Task fallback changes discard incompatible stale temperature settings
* Agent phase timings remain monotonic and consistent

## Current documentation

[Evals](/docs/production/evals) · [Observability](/docs/production/observability) · [Model governance](/docs/production/model-governance) · [Run agents with files](/docs/api/run-agents)