Service
AI agent development on Google Cloud
Formare AI designs, builds and deploys custom AI agents on Google Cloud: single agents with 5–10 tools, multi-agent systems on Google's Agent Development Kit (ADK) and Gemini, and the Cloud Run deployment, session storage, audit logging and security around them. A single agent with 5–10 tools typically takes 2–3 weeks; a full multi-agent system with a React frontend and production infrastructure takes 6–10 weeks. Work is priced at €600–€900 per day or from €8,000 for a fixed-scope project, and ends with documentation, runbooks, cloud credentials and 30 days of email support.
Updated
Scope
What we build
An agent is a model that decides which tools to call, and in what order, to finish a task: look up a record, call an API, write to a database, hand work to another agent. We build the agent, the tools it calls and the service it runs in.
- Gemini agentic loops with function calling, such as a booking agent with 11 tools for geocoding, fares, conflict-safe booking, payments, calendar and flight tracking
- Multi-agent systems on Google ADK: an orchestrator that classifies intent, splits the work and routes it to specialist agents running in parallel
- Agents over your data, such as natural-language-to-SQL on BigQuery that is schema-aware, so it cannot invent column names
- Interoperability through the A2A (Agent-to-Agent) protocol, and MCP servers that expose your systems to any MCP-compatible assistant
- The product around the agent: FastAPI or Express backend, React frontend, PostgreSQL, authentication and payments, deployed on Cloud Run
Production
Why agents fail between prototype and production
An agent that works in a demo usually has a simple tool loop and an in-process session store. The failures below are the ones our production guide describes, and none of them is about model quality.
What breaks
- Tool calls that time out silently, so the agent reports success on stale data
- Session state held in memory and lost on every Cloud Run restart
- No distributed tracing, so nobody can tell which sub-agent produced a wrong answer
- Token costs that grow with conversation length
- No circuit breakers on external APIs, so one slow dependency stalls the whole agent graph
What goes in from the first sprint
- Session persistence in Cloud SQL through ADK's DatabaseSessionService instead of in-memory sessions
- A timeout, retry with backoff and circuit breaker around every external tool
- An audit log of every tool call and result in BigQuery, plus tracing in Cloud Trace
- Caching of tool results and model context to keep token spend predictable
- A Cloud Run deployment in the first sprint, so testing happens on real infrastructure
Cost and timeline
How long an AI agent project takes and what it costs
A single AI agent with 5–10 tools and Cloud Run deployment typically takes 2–3 weeks. A full multi-agent system with a React frontend and production infrastructure takes 6–10 weeks. Most engagements run 2–12 weeks, and scope is defined before work begins.
Pricing is €600–€900 per day, from €8,000 for a fixed-scope project paid in milestones, or from €4,500/mo for a retainer. Rates are in euros and VAT may apply. If you are not sure whether you need agents, a data pipeline or just a cleaner data model, the 4-Week AI Strategy & Architecture Audit (from €8,000) delivers a written architecture plan prioritised by ROI. The 30-minute discovery call is free.
Work
Agent systems we have built
AI Airport Transfer Platform
George Travel — georgetravel.co.uk
Full-stack AI booking platform for a UK private driver. Gemini agentic loop with 11 custom tools handles geocoding, fare calculation, conflict-safe booking (PostgreSQL advisory locks), Stripe payments, Google Calendar, automated emails, real-time flight tracking, and weather. Interoperable via A2A protocol and MCP.
- 11 AI tools
- 0 double-bookings since launch
- A2A + MCP interop
Senior Merchandiser AI Agent
Retail catalog intelligence — multi-agent
Multi-agent merchandiser: Gemini planner, parallel vector search and re-ranking, Feature Store enrichment.
- +15–25% search precision
- Real-time Feature Store <2ms
NL → BigQuery Assistant
Conversational data analysis — multi-agent
Ask BigQuery in plain English: an orchestrator routes to schema-aware NL→SQL and Pandas agents.
Personal AI Assistant — Bucharest
Romanian market — restaurant + reservations
Bucharest restaurant discovery with Google Search grounding, and reservations through a custom MCP tool.
FAQ
Questions about AI agent development
Cost, timelines and fit, answered before the first call.
What is the difference between an AI agent and a chatbot?
A chatbot generates replies. An AI agent also acts: it decides which tools to call, such as a database query, a payment API or a calendar, runs them in a loop and uses the results to finish a task. That makes reliability engineering (timeouts, retries, session state, audit logs) as important as the model itself.
Which framework and cloud do you build AI agents on?
Google's Agent Development Kit (ADK) with Gemini models on Vertex AI, served from FastAPI on Cloud Run. Agents interoperate through the A2A protocol and MCP servers when a project needs it. The discovery call also covers whether Google Cloud is the right platform for your use case at all.
How long does it take to build and deploy a production AI agent?
A single AI agent with 5–10 tools and Cloud Run deployment is typically 2–3 weeks. A full multi-agent system with a React frontend and production infrastructure runs 6–10 weeks. Most projects run 2–12 weeks, and scope is defined before any work begins.
How much does a custom AI agent cost?
Pricing is €600–€900 per day, from €8,000 for a fixed-scope project paid in milestones, or from €4,500/mo for a retainer. Rates are in euros and VAT may apply. If the right architecture is not yet clear, the 4-Week AI Strategy & Architecture Audit (from €8,000) ends in a written plan before you commit to a build. The 30-minute discovery call is free.
Can the agents work with our existing systems and data?
Yes. Agents reach your systems through custom tools and MCP servers, and we regularly integrate with existing BigQuery datasets, Cloud Run services, Vertex AI resources and GCP IAM configurations. The discovery call is specifically for understanding your current stack before scoping begins.
How do you handle GDPR and the EU AI Act for AI agents?
We design for EU-region data residency, audit trails and explainability from the first sprint rather than after launch. The architecture and code-level controls are set out in our article on EU AI Act and GDPR compliance for RAG systems on Google Cloud. That is engineering guidance, not legal advice, so your counsel should confirm how the AI Act classifies your use case.