Philipp Schmid is a Staff Engineer at Google DeepMind building the organization’s AI developer-experience and developer-relations team. He helps developers turn Gemini, Gemma, and other advanced models into usable software, with particular attention to agent reliability, practical tooling, and evaluation.
He joined Google DeepMind in 2025. His current work encompasses Gemini and Gemma across cloud and on-device environments, combining developer education with the engineering questions involved in building useful agents: conversation state, multimodal inputs, tool integration, inference costs, and reliable execution.
Context is application state. Agents operate on user intent, preferences, previous decisions, and changing instructions that cannot always be reduced to predefined flags. Schmid argues that developers should specify goals and constraints while allowing models latitude in how they achieve them. His account of why experienced engineers struggle with agents emphasizes preserving semantic context and designing tools whose descriptions expose assumptions human programmers might leave implicit.
Reliability requires repeated evaluation. A single successful demonstration cannot establish whether a nondeterministic agent works consistently. Schmid favors repeated trials, execution traces, outcome-based grading, and human or model-assisted assessment. His analysis of multi-turn coding evaluation extends that concern to changing requirements and earlier architectural decisions that can compromise later work.
Skills need evals and an expiration date. Schmid distinguishes temporary capability skills, which compensate for current model limitations, from durable preference skills encoding organization-specific practices. He recommends precise triggering instructions, negative test cases, progressive disclosure of references, and with-and-without-skill ablation tests. His Gemini Interactions API evaluation example demonstrates how trace inspection and lightweight validators can establish whether specialized guidance improves performance or should be retired.
Errors are inputs, not dead ends. Failed tool calls should inform an agent’s next action instead of automatically restarting lengthy workflows and discarding accumulated context. Schmid also treats interoperability pragmatically: the Model Context Protocol can reduce duplicated integration work, but models still select tools through ordinary function calling. His public analysis of multi-agent handoffs similarly concentrates on explicit coordination and stateless interaction.
A hands-on path through Gemini’s Python SDK, token accounting, document extraction, native tools and MCP, with the state, cost and execution boundaries made explicit.
Build a small coding agent, give a conversational model tools and vision, and follow the deployment and reliability tradeoffs exposed by the live demonstrations.
A skill is useful only if it changes an agent’s behavior for the better. Small, repeatable evaluations reveal missed triggers, regressions, wasted context and when a skill can retire.
Building reliable agents means preserving intent, letting plans change, recovering from failures, and evaluating outcomes instead of prescribing every step.
Philipp Schmid rebuilds a GitHub pull request reviewer three times, moving from a handwritten Python loop to a hosted agent with Bash, files, and search. The progression shows which code can disappear, where execution moves, and why instructions and evaluations remain your responsibility.
Frameworks can remove handwritten loops and schemas while leaving the developer responsible for tool implementations and hosting.
Skills and saved files can extend capabilities or carry preferences and handoffs. Improved models should create opportunities to delete orchestration that no longer earns its place.
▶ Watch ↗
