Gemini Managed Agents QA workflows need a fresh review after Google announced new Gemini API Managed Agents capabilities on July 7, 2026. The update adds background execution, remote MCP server integration, custom function calling, and credential refresh across interactions.
Google positions the change as a step toward more reliable production agents. For QA engineers and automation teams, the practical question is narrower: how do you validate an agent that can keep working asynchronously, call external tools, and preserve a remote sandbox state across interactions?
What Google announced
- Background execution: Google says developers can run interactions asynchronously on the server and use the returned ID to poll, stream progress, or reconnect later.
- Remote MCP: Managed agents can connect directly to remote Model Context Protocol servers, letting agents use external tools and APIs from the sandbox.
- Custom function calling: Built-in sandbox tools can run on the server while client-owned business logic can move the interaction into a required-action step.
- Credential refresh: Google says developers can refresh network credentials on a later interaction while keeping sandbox files, installed packages, and cloned repositories intact.
Why this matters for QA engineers
This update moves agent testing closer to real application testing. QA teams may now need to validate behavior that happens after the user disconnects, after credentials rotate, or after an agent calls a remote tool. A passing prompt response is not enough evidence.
- Long-running tasks need status checks: test queued, running, completed, failed, canceled, and resumed states instead of only testing the first request.
- MCP tool access expands risk: verify authentication, authorization, audit logs, prompt-injection handling, and least-privilege tool descriptions before agents touch internal systems.
- Credential refresh needs regression coverage: test expired tokens, rotated keys, revoked access, and sandbox state preservation.
- Agent evidence must be reviewable: capture interaction IDs, tool calls, status transitions, terminal output, and final artifacts so failures can be reproduced.
A QA checklist for the update
- Run one background task that completes normally and one that intentionally fails.
- Disconnect the client and verify polling or reconnect behavior returns the correct final state.
- Connect a test MCP server with harmless tools first, then confirm the agent cannot call tools outside its allowed scope.
- Rotate a short-lived credential during a multi-step interaction and verify both access control and sandbox continuity.
- Review logs for every external tool call, including parameters, response status, and the user-visible outcome.
Bottom line
Google’s Gemini API Managed Agents update is meaningful because it adds operational features, not just model features. QA teams evaluating AI agents should treat background execution, remote MCP, and credential refresh as testable product surfaces. Start with small sandboxed workflows, require logs and artifacts, and do not connect agents to sensitive QA or production systems until access boundaries are proven.
Sources
- Google Blog: Expanding Managed Agents in Gemini API – July 7, 2026.
- Google AI for Developers: Antigravity Agent documentation – accessed July 8, 2026.