On May 20, 2026, alongside Google I/O 2026, articles were published announcing that Google Unveils Antigravity 2.0 and Gemini API Managed Agents and Google Unveils "Managed Agent API" to Launch AI Agents with Hosted Linux Environments via Single API Call and Custom Markdown Instructions. Entering an era where an AI agent equipped with a Linux environment spins up with a single API call, AI Studio gained official support for Android app development, while Antigravity 2.0 showcased a demo developing an operating system from scratch and running Doom on it.
For firms driving AI adoption for mid-sized enterprises through client development, this marks a turning point where all clients in the PoC phase face the simultaneous decision of building an agent execution infrastructure on self-hosted IaaS or adopting Google-hosted Managed Agents. Following Anthropic-based managed agents covered in our article on Claude Managed Agents, the same concept has now been standardized within the Google ecosystem. In this article, we outline key architectural considerations for integrating and operating Managed Agents in business workflows from a client service perspective.
Why the Managed Agents API transforms architectures for mid-sized enterprises
| Item | In-house build (traditional) | Google Managed Agents |
|---|---|---|
| Execution environment | Custom in-house builds using GKE, Cloud Run, or Lambda | Linux environment launched via a single API call |
| Agent definition | Code + IaC + deployments | Custom instructions in Markdown |
| Scaling | Custom monitoring + auto-scaling architecture | Transparently managed by Google |
| Tool connectivity | Custom API integration and MCP implementation | Google Connectors + standard MCP |
| Audit Logging | Custom SIEM integration | Standard Cloud Audit Logs provided |
| Model switching | Code releases | API parameter updates only |
In short, Managed Agents represent the transition of operational agent IaaS into SaaS. At the same time, defining business scopes, data boundaries, and human approval gates remains an essential design responsibility in client development.
Three structures transformed by Managed Agents
Structure 1: From implementation to architecture and operations
Previously, the primary focus when building operational business agents for clients was constructing the execution infrastructure. With the rise of Managed Agents, the center of gravity shifts to defining business scope, designing prompts, establishing human approval gates, and managing audit operations. The proportion of engineering hours decreases, while the emphasis on business analysis, legal compliance, and operations increases.
Structure 2: From a single cloud to Google × MCP × line-of-business systems
Managed Agents reach line-of-business systems outside Google Workspace via MCP and Google Connectors. In custom development engagements, we combine this with transforming internal systems into agents via MCP Tunnels to design communication paths governed by Zero Trust and least privilege.
Structure 3: From multi-year IaC investments to agent definitions in Markdown
Because agent behavior can now be defined using custom instructions in Markdown, iteration cycles with business units are shortened dramatically. At the same time, this creates a new governance challenge: version controlling business logic defined in Markdown.
Five phases of integrating Managed Agents into business operations
Phase 1: Defining business scope and aligning expectations (2 weeks)
We audit which workflows, with which data, and to what degree to automate across three dimensions: operations, data, and legal compliance. Workflows involving personal data or confidential classifications require mandatory approval gates.
Phase 2: Managed Agents PoC (2–3 weeks)
- Initial setup of Gemini API and Managed Agents
- Prototyping 3–5 business scenarios
- Designing and reviewing Markdown custom instructions
- Connecting line-of-business systems (MCP / Connectors)
- Failure fallbacks and human approval gates
Phase 3: Production rollout (3–4 weeks)
- Cloud IAM and confidentiality label integration
- Integrating audit logs (Cloud Audit Logs) with SIEM
- DLP and guardrails (architected similarly to our AI Agent Production DB Deletion Guardrails)
- Phased rollout (by organizational unit → department → company-wide)
- Establishing operational runbooks
Phase 4: Internal training and knowledge base setup (2 weeks)
- "Authoring Agents in Markdown" training for business units
- Usage guidelines, anti-patterns, and escalation paths
- Management training for department leads
Phase 5: Monthly operational reviews (ongoing)
- Audit log reviews and anomaly detection
- Refining agent behavior (prompts and tool integrations)
- KPI monitoring (processed volume, time saved, error rate)
- Tracking new capabilities (Gemini 3.5 and Antigravity 2.0 updates)
- Version control reviews for Markdown instructions
Standard technology stack set for custom development
| Layer | Recommended technology | Alternative |
|---|---|---|
| Agent foundation | Gemini API Managed Agents | Anthropic Managed Agents |
| Model | Gemini 3.5 Pro / Ultra | Gemini 2.5 Flash (lightweight tasks) |
| Tool connectivity | MCP + Google Connectors | Custom APIs |
| IDE and development support | Antigravity 2.0 | Cursor / Cline |
| Authentication / authorization | Google Workspace IdP + IAM | Individual OAuth implementations |
| Audit Logging | Cloud Audit Log + BigQuery + Looker | Splunk |
| DLP and data governance | Workspace DLP + sensitivity labels | Microsoft Purview |
| Approval workflow | Apps Script + Slack | ServiceNow |
Which projects are suitable (and unsuitable) for Managed Agents
| Suited projects | Unsuitable projects |
|---|---|
| Company-wide Google Workspace adoption | Primarily Microsoft 365 |
| Multiple line-of-business system connections that need consolidation via MCP | Contained within a single SaaS application |
| Desire to centralize audit logging | Existing company-mandated logging infrastructure |
| Need for rapid prototyping cycles with business departments | Waterfall development with frozen specifications |
| Desire to reduce the cost of building custom agent execution infrastructure | Existing in-house Kubernetes infrastructure |
Six clauses to include in client contracts
| Clause | Details | What the client should verify |
|---|---|---|
| Target operations | Department × function × sensitivity tier | Out-of-scope items require separate agreements |
| Responsibility for versioning Markdown instructions | Repositories, review workflows, and release procedures | Executive approval |
| Incident SLA | Response times for erroneous executions or suspected leaks | Business impact level |
| Audit log retention | 36+ months / BigQuery storage | Legal review |
| Offboarding handover | Configurations + instruction documents + audit results | Internal operational continuity |
| Handling of personal data | Data scope and third-party disclosure | PPC notification filings |
Client ROI estimate (assuming 200 employees and 10 operational agents)
| Item | In-house build | Google Managed Agents custom development engagement | Difference |
|---|---|---|---|
| Initial setup timeframe | 6 months | 6–8 weeks | -4 months |
| Monthly operational workload | 200h | 30h | -170h |
| Monthly working hours saved | Average of 6 hours / person | Average of 14 hours / person | +8h |
| 200 people × 8 hours × 12 months | — | — | +19,200h |
| Audit response hours | 300 hours/year | 80h/year | -220h |
| Annual benefit | — | — | Equivalent to approximately ¥150 million |
This translates to over ¥120 million annually, calculated at an hourly rate of ¥6,000. Compared to building and maintaining a custom execution platform in-house, the reduction in operational overhead and audit compliance workload serves as the primary basis for investment decisions.
Five common pitfalls
Pitfall 1: Misunderstanding that "written in Markdown" means anyone can build it
While writing Markdown is straightforward, failing to design business scopes, data boundaries, and human approval gates directly leads to production incidents. Agent design should involve three roles: business PM, engineer, and legal counsel.
Pitfall 2: Promoting a PoC straight into production
PoCs operate under the assumption of using mock or sandbox data. Moving to production requires a comprehensive redesign of permission boundaries, audit logging, and fail-safes.
Pitfall 3: Lacking version control for Markdown
If Markdown instructions are edited directly in Google Drive or Notion, tracking who changed what becomes impossible. Manage them using Git repositories and pull request reviews.
Pitfall 4: Behavioral shifts caused by model updates (Gemini 3.5 → 4.0)
While Managed Agents make switching models effortless, they also introduce the risk of unannounced behavioral changes. Conduct monthly regression testing on key operational scenarios.
Pitfall 5: Collecting audit logs without active analysis
Simply enabling Cloud Audit Logs does not provide anomaly detection. Build operations that include dashboards in BigQuery and Looker alongside monthly review meetings.
90-day action plan
| Week | Action |
|---|---|
| Week 1〜2 | Inventorying operations, data, and legal requirements |
| Week 3〜5 | Managed Agents PoC (3–5 business scenarios) |
| Week 6〜9 | Production deployment (IAM, auditing, DLP, guardrails) |
| Week 10〜11 | Internal training + phased rollout |
| Week 12〜13 | Launching monthly operations + initiating KPI tracking |
Summary — Client architecture design in the era when operational agent IaaS becomes SaaS
The Google Managed Agents API and Antigravity 2.0 have established a viable path to shift operational agent execution infrastructure from custom in-house builds to Google-hosted services. For partners driving enterprise AI adoption, integrating business scoping, Markdown instruction version control, and audit operations delivers far greater strategic value than the raw code implementation itself.
Integrating Managed Agents into business workflows requires tailored architectural design depending on workflow scope, data confidentiality tiers, and the number of connected systems; we quote each project individually. Whether you are facing challenges where you want to trial Managed Agents but legal has stalled progress, you built a custom Kubernetes agent infrastructure but cannot manage operations, or version control for Markdown instructions has broken down, feel free to contact us via our contact form.
Sources
- Google Announces Antigravity 2.0 and Gemini API Managed Agents (gihyo.jp)
- Launch AI Agents with Hosted Linux Environments via Single API Call: Google Announces "Managed Agent API" Supporting Custom Markdown Instructions (Publickey)
- Google Announces Antigravity 2.0 (Publickey)
- Claude Managed Agents Client Services (GH Media)
- Custom Integration of Internal Systems as Agents via MCP Tunnels (GH Media)
- AI Agent Production DB Deletion Guardrails (GH Media)
