About
Member of Technical Staff / Technical Product Manager at Google, focused on agent security, governance, capabilities, quality, and evaluation
About Alan Blount
I’m a Member of Technical Staff / Technical Product Manager at Google in Boulder, Colorado, working on the Gemini Enterprise Agent Platform (formerly Vertex AI). I spent about 20 years as a software engineer before moving into product. I help developers build agents they can trust in production.
My day-to-day covers A2A, ADK, Agent Engine, MCP, grounding/RAG, memory, and Agent Ops. Before that I worked on Google’s Payments Smart Router, an ML multi-armed-bandit optimization and analytics system. Before Google, I was a tech lead and full-stack developer, helping startups grow into successful small companies since ~2001.
Areas of Focus
Security & Governance
Agents are a new kind of principal. To be useful they need autonomy and tools, and every capability you give them adds risk. I work on making that trade-off explicit and enforceable:
- Agent identity: each agent gets its own verifiable identity (e.g. SPIFFE), separate from the user who invoked it and the developer who built it, with least-privilege permissions so a compromised agent has a small blast radius.
- Policy & authorization: constraining access to tools, data, sub-agents, and remote agents, enforced at the API/gateway layer for MCP and A2A traffic.
- Defense in depth: deterministic guardrails outside the model, plus reasoning-based defenses (guard models, plan review) against prompt injection, data poisoning, and data leakage.
- Governance at fleet scale: a central control plane instead of “agent sprawl”. Every user→agent, agent→tool, agent→agent, and agent→model call is authenticated, authorized, logged, and auditable.
- Execution isolation: sandboxes for agent-generated code, where network egress is often a bigger attack surface than the hypervisor.
Capabilities, Quality & Evaluation
Agents are probabilistic by design, so assert output == expected isn’t enough. I care about measuring whether agents (and software in general) actually work:
- Capabilities: knowing what each level of agent can do, from a single connected problem-solver to multi-agent systems and self-evolving agents, and scoping systems to match.
- Agent Ops: the next step after DevOps and MLOps. Metrics-driven development, go/no-go deployment gates, and A/B-style KPIs tied to real outcomes (goal completion, cost per task, latency, user satisfaction).
- Evaluation: LM-as-judge quality scoring, trajectory evals, and human feedback that feeds back into automated test sets.
- Observability: OpenTelemetry traces that answer why an agent did what it did.
- Deterministic verification first: wherever possible, check results with code instead of opinion.
- Tokenomics: routing work between fast and deep models by task ROI, not habit.
Writing & Publications
- Introduction to Agents whitepaper (lead author, with Antonio Gulli, Shubham Saboo, Michael Zimmermann, and Vladimir Vuskovic). Day 1 of the Kaggle × Google 5-Day AI Agents Intensive (Nov 2025, 1.5M+ learners). It covers agent taxonomy, architecture, Agent Ops, agent identity, policy, and governance. I was also a guest on the Day 1 livestream.
- The full 5-day series:
- Google Developers Blog articles on ADK, A2A, MCP, AG-UI, and Agent Plugins
- Agentic AI Foundation (AAIF) blog
- Google Cloud Tech articles on X: agent sandboxes (with Ryan Ismert) and Fable 5.1 + Gemini 3.8 Flash model routing
Talks
- Agents and MCP @ Google Scale, MCP Dev Summit NA 2026 (with Vaibhav Katkade)
- How Google Built the A2A Protocol, Patchwork: Syncing on AI Standards (2026)
Open Source & Community
- Creator of Headroom: context compression for AI applications that reduces token costs by 70%
- Agent protocols & runtimes: A2A, ADK, MCP, plus OpenClaw, Scion, Flue, and ax
- zeroasterisk on GitHub: developer tools and AI-related projects
- Mentoring: free coding bootcamps, universities, conferences, and meetups
Technical Philosophy
- Research-driven development: understand the problem before building solutions
- Iterative prototyping: quick feedback loops and continuous improvement
- Deterministic verification: prefer programmatic checks over subjective evaluation
- Platform-native solutions: build with the grain of existing systems
- Trust is the product: building an agent is easy; trusting it is hard
Personal Side
When I’m not working on AI systems, you’ll find me:
- Parenting: raising kids with my wife Anita in Boulder
- Teaching: helping kids balance screen time with productive projects
- Learning: always exploring new technologies and methodologies
- Writing: documenting experiences both technical and personal
Boulder, Colorado
I love living in Boulder. It’s the perfect blend of tech innovation and outdoor adventure, with incredible access to hiking, biking, and the mountains, plus a thriving tech community.
Connect
I’m always interested in discussing:
- Agent security, identity, and governance
- Agent quality, evaluation, and Agent Ops
- Agent interoperability (A2A, MCP) and memory
- The intersection of technology and parenting