In a pointed assessment with direct guidance, Google leader Antonio Gulli set out why many AI agents fail after launch and which patterns can keep them working. Speaking this week, he framed the issue as a design and deployment gap that shows up once systems meet real users and messy data.
Google’s Antonio Gulli explains why most AI agents break in production—and the patterns needed to fix them.
His message comes as companies rush to embed AI-driven agents into customer support, software operations, and business workflows. Many teams report early wins in tests, only to see brittle behavior, higher costs, and user frustration in real use. Gulli’s focus on repeatable patterns aligns with a push inside the field to make these systems more reliable, observable, and safe.
Why AI Agents Fail After Launch
AI agents often work in demos but stumble with real traffic. Inputs grow noisy. Edge cases pile up. Integrations drift as APIs change. Guardrails miss new prompts. Small errors stack into larger failures.
Teams also struggle with monitoring. Many track only output quality, not the steps that lead to it. Without traces, it is hard to find the cause of a sudden drop in accuracy or a spike in latency. Costs can climb as agents loop or retry without bounds.
Organizational pressure adds risk. Deadlines shorten testing cycles. Upstream data shifts faster than agents adapt. In production, this mix can cause silent failures that are hard to diagnose.
Patterns That Improve Reliability
Gulli’s focus on patterns points to engineering practices that make agents sturdier and easier to manage. The most cited approaches in the field include process checks, clear interfaces, and runtime controls designed for real-world use.
- Plan-and-Execute Loops: Separate planning from action. Force the agent to write a plan, then follow it step by step with checks.
- Tool Contracts: Define strict input and output formats for tools and APIs. Validate before calls and after responses.
- Grounding and Retrieval: Use retrieval to anchor outputs in current data. Log sources so answers can be audited.
- Deterministic Fallbacks: Provide safe, simple paths when confidence is low or timeouts hit.
- Guardrails and Policies: Enforce rules on inputs, actions, and outputs. Block unsafe steps before they happen.
- Observability by Design: Capture traces of every decision, tool call, token budget, and user signal for later review.
- Evaluation Pipelines: Run offline suites and shadow tests that mirror production traffic and edge cases.
These patterns do not remove model errors, but they change failure from chaotic to predictable. That shift lets teams fix issues without full rebuilds.
From Pilot To Production
The jump from a lab demo to a live system is often where agents falter. Controlled tests rarely reflect real user intent. Integration errors hide until the first traffic surge. Teams need staging runs that mirror production data, load, and failure modes.
Versioning is key. Track model versions, prompts, tools, and policies as one release. Tie live metrics to each version. If quality drops, roll back fast. Treat prompts like code with reviews and automated tests.
Human feedback loops matter. Route unclear cases to reviewers. Use their notes to improve prompts, tools, and rules. Over time, the share of auto-resolved cases should rise while risk stays within set limits.
Business Impact And Risk
For many companies, the main risks are reputational harm, runaway spend, and compliance gaps. A failed agent can give wrong answers, leak data, or take the wrong action. Costs can spike when loops or retries go unchecked.
Clear service levels help. Define targets for accuracy, latency, and containment rate. Align these with cost budgets and safety rules. Make trade-offs explicit so product and risk teams can agree on what “good” looks like.
Sectors with strict rules—health, finance, and government—will demand tighter controls. Audit trails, red-teaming, and policy checks must be built in from the start.
What Comes Next
Gulli’s focus on repeatable patterns reflects a maturing phase for AI agents. The field is moving from quick demos to durable services. That shift favors teams that invest in testing, tracing, and policy engineering.
Companies that adopt these practices can cut failures, steady costs, and earn user trust. Those that skip them will keep reliving the same outages and incident reviews.
The key takeaway is simple. Agents fail less when they follow clear plans, use reliable tools, and operate under steady oversight. Expect more playbooks and shared benchmarks as teams learn from each other and turn these patterns into standard practice.
Deanna Ritchie is a managing editor at DevX. She has a degree in English Literature. She has written 2000+ articles on getting out of debt and mastering your finances. She has edited over 60,000 articles in her life. She has a passion for helping writers inspire others through their words. Deanna has also been an editor at Entrepreneur Magazine and ReadWrite.
