Google Managed Agents Gemini Api Developer Guide 2026
Date: 2026-05-25 Content track: sandbox-poc Slug: google-managed-agents-gemini-api-developer-guide-2026
Purpose
Evaluate the developer preparation pattern behind Google Managed Agents without claiming live Gemini API execution. The sandbox checks whether an AGENTS.md plus SKILL.md package can express role, constraints, and a narrow task process, then simulates a policy gate around allowed and denied actions.
Environment
- Host: local macOS workspace
- Sandbox directory:
/tmp/effloow-google-managed-agent-poc - Runtime: Node.js v25.9.0
- Network/API calls: none
- Secrets used: none
- Production files touched: no production data, no
.env
Commands
rm -rf /tmp/effloow-google-managed-agent-poc
mkdir -p /tmp/effloow-google-managed-agent-poc/skills/invoice-audit
cd /tmp/effloow-google-managed-agent-poc
node --version
node validate-agent-package.mjs
node simulate-policy-gate.mjs
sed -n '1,80p' artifacts/report.md
Generated Sandbox Files
AGENTS.md contained:
# Invoice Evidence Agent
Role: inspect uploaded invoice CSV files and produce a reconciliation checklist.
Constraints:
- Never call external payment APIs.
- Never mutate source files.
- Write findings to artifacts/report.md.
- Ask for human approval before any irreversible action.
skills/invoice-audit/SKILL.md contained:
# invoice-audit
Use this skill when a CSV contains invoice_id, amount, currency, and status columns.
Process:
1. Parse CSV headers.
2. Count rows by status.
3. Flag negative amounts and missing currencies.
4. Emit a markdown summary.
Output
node --version
v25.9.0
node validate-agent-package.mjs
{
"hasRole": true,
"hasHumanApprovalRule": true,
"hasNoMutationRule": true,
"hasSkillProcess": true
}
node simulate-policy-gate.mjs
{
"summary": {
"totalRows": 4,
"byStatus": {
"paid": 2,
"pending": 1,
"failed": 1
},
"negativeAmounts": [
"INV-101"
],
"missingCurrency": [
"INV-102"
]
},
"allowed": [
"read_csv",
"write_report"
],
"denied": [
"refund_payment"
],
"report": "artifacts/report.md"
}
artifacts/report.md
{
"summary": {
"totalRows": 4,
"byStatus": {
"paid": 2,
"pending": 1,
"failed": 1
},
"negativeAmounts": [
"INV-101"
],
"missingCurrency": [
"INV-102"
]
},
"allowed": [
"read_csv",
"write_report"
],
"denied": [
"refund_payment"
]
}
What Worked
- A small
AGENTS.mdfile was enough to encode role, no-mutation, no-external-payment, and human-approval rules. - A separate
SKILL.mdfile cleanly described the task-specific invoice-audit process. - The local validator caught whether the package included the minimum role and safety constraints.
- The policy simulation allowed read/report actions and denied the irreversible
refund_paymentaction. - The run produced an artifact file, which mirrors the operational expectation that agent work should leave inspectable outputs.
What Failed Or Was Not Tested
- No live Gemini API call was made because this article workflow must not use production credentials and no API key was provided for this isolated run.
- The PoC did not verify Google-hosted sandbox lifecycle behavior, billing, region availability, quota behavior, or preview access.
- The PoC did not test MCP server connectivity, GitHub repository mounting, Google Cloud Storage sources, or workspace download APIs.
- The policy gate is a local deterministic simulation, not Google's managed-agent authorization layer.
Sources Checked
- Google announcement: https://blog.google/innovation-and-ai/technology/developers-tools/managed-agents-gemini-api/
- Gemini Interactions API docs: https://ai.google.dev/gemini-api/docs/interactions
- Gemini tools docs: https://ai.google.dev/gemini-api/docs/tools
- Gemini CLI subagents documentation: https://github.com/google-gemini/gemini-cli/blob/main/docs/core/subagents.md
- DataCamp Managed Agents tutorial: https://www.datacamp.com/tutorial/managed-agents-in-the-gemini-api
- GIGAZINE coverage: https://gigazine.net/gsc_news/en/20260520-google-gemini-api-managed-agents/
- VentureBeat analysis: https://venturebeat.com/orchestration/googles-managed-agents-api-promises-one-call-deployment-at-the-cost-of-execution-layer-control/
Article Boundary
The article may truthfully say Effloow Lab ran a local sandbox PoC for agent-package validation and policy-gate simulation. It must not say Effloow Lab ran Gemini Managed Agents through the live API, verified Google's sandbox, measured performance, validated pricing, or tested MCP connectivity.
Read the article
This note supports the public article and records what was actually checked.