AI DevCamp

# Build

Sonika Janagill

Hands-on Build workshop with Sonika Janagill. Set up Google Cloud (new project, free credits, billing alerts, and a Vertex AI spend cap), then local tools: gcloud CLI, uv, Python, Node, compute region us-central1, and model location global. Build an ADK agent with agent CLI and the playground. The running example is a social-post agent: a root orchestrator, a research agent (Google Search), a draft agent, Agent Skills (brand voice, platform limits, post format), image generation, and human confirmation before anything is published. Buffer MCP posts to X or LinkedIn; a small custom MCP is the optional LinkedIn path. The web UI is CopilotKit (AG-UI) on localhost:3000, talking to the agent as an API. You leave with a local ADK app that can research, draft, and post only after you approve. Live audio/video is a stretch (pinned live model, region us-central1, higher cost — test in text first). Evals and Agent Platform deploy come in later weeks.

# Scale

Sonika Janagill

Scale theory with Sonika Janagill. Scale means many people can use the agent at once without it crashing or running out of resources — off a laptop, with sessions, identity, and logging. Recap of Build: ADK (model, instructions, MCP tools, Agent Skills), a multi-agent orchestrator for research, drafting, and posting, and a CopilotKit UI on AG-UI. Then Agent Platform's scale stage: Agent Runtime (warm start under a second, long-running work up to 7 days, up to 3,000 agents per project, five deploy paths), sessions versus Memory Bank (embedding search, 90-day topic TTL, separate banks per agent and user), code sandboxes including computer use, and Agent Identity so actions are not only attributed to a person. The web UI stays private on Cloud Run behind Identity-Aware Proxy; the runtime API is called with bearer tokens. Do not hard-code service accounts. Saturday's workshop is deploying the app you already run locally. Weekly submission is the lab plus your own parallel agent.

# Scale

Sonika Janagill

Saturday Scale lab with Sonika Janagill. Build a multi-agent app in Agent Studio (no code required), then copy the generated ADK project and deploy it with the agent CLI. The deploy uses a manifest and an Agent Identity (a cryptographic principal, not a shared service account). After deploy: call the runtime with curl or Postman (OpenAPI, bearer token), turn on sessions so conversations and traces stay in the console, and attach a Memory Bank scoped per agent and per user. A Cloud Run front end connects with a backend-origin URL; gcloud run services proxy reaches a private service on a local port. FastAPI in main.py wraps the agent when you need your own API. Agent Studio is text-only today. Live audio and video stay on local ADK. Vector Search syncs indexes for you; a RAG corpus does not — re-sync files yourself. Week 2 submission is two deploys: your agent and the social poster (screenshot or GitHub link on Discord). Agent Runtime is on a free tier through the end of 2026; delete Cloud Run services and agents you are not using.

# Govern

Renuka Kelkar

Threat model for this agent (injection, PII, off-brand drafts, posting without approval). Defence in depth: input screening, tool-call policy, output screening, human approval. Governance stack: Identity, Registry, Gateway, Policies. Model Armor in ADK, Cloud DLP, HITL as governance, Skills as policy. Compliance corner. Live demo: injection blocked, PII redacted — then Saturday's plan.

# Govern

Renuka Kelkar

Codelab: add Model Armor checks to the agent pipeline. Build an evaluation set for the drafting agent. Harden the WhatsApp/Discord approval step (reject, timeout, edit). Test against adversarial inputs. You leave with a governed agent: screening, PII redaction, policy, registry — and a survived red-team round.

# Optimise

Nishi Ajmera

Evals as behaviour tests (trajectory vs response, rubrics, golden set). Multi-Turn AutoRaters and online evaluation. Simulation. Observability: Cloud Trace, Unified Trace Viewer, BigQuery Agent Analytics. Prompt optimisation from failure patterns. State and memory as token levers. Cost: tokens, tool calls, runtime; model right-sizing. Evals in the pipeline: push → eval gate → deploy. Live demo, Saturday plan, and graduation slots. Assignment and show-and-tell framing.

# Optimise

Nishi Ajmera · Sumith Damodaran · Renuka Kelkar · Sonika Janagill

Apply cost optimisation (right-size models per sub-agent). Set up a basic CI/CD pipeline with an eval gate. Deploy the final version. Live demos: full agent run end to end — audio/text input to an approved, posted LinkedIn/X post. Graduation demos. You leave with a measured agent: golden eval set, traces, analytics, eval-gated deploy.