The enterprise AI market is currently suffering from a huge hangover. Over the past two years, decision makers have been inundated with demonstrations of autonomous agents booking flights, writing code, and analyzing data. However, the reality on the ground is totally different. While experimentation is at an all-time high, deploying reliable, autonomous agents in production remains a challenge.
A recent study by MIT’s NANDA Project highlighted a sobering statistic: approximately 95% of AI projects fail to deliver ultimate value. They hit walls when moving from the sandbox to the real world, often breaking under the weight of edge cases, hallucinations, or integration failures.
According to Antonio Gulli, a senior engineer at Google and director of the CTO engineering office, the industry suffers from a fundamental misunderstanding about what agents really are. We’ve treated them like magic boxes rather than complex software systems. "AI engineering, especially with large models and agents, is really no different from any form of engineering, such as software or civil engineering." Gulli said in an exclusive interview with VentureBeat. "To build something lasting, you can’t just chase the latest model or framework."
Gulli maintains that the solution to "valley of disappointment" It is not a smarter model, but a better architecture. His recent book, "Agent Design Patterns," provides rigorous and repeatable architectural standards that make "toy" agents in reliable business tools. The book pays homage to the original. "Design patterns" (one of my favorite books on software engineering), which brought order to object-oriented programming in the 1990s.
Gulli presents 21 fundamental patterns that serve as the basis for reliable agent systems. These are practical engineering structures that dictate how an agent thinks, remembers, and acts. "Of course, it is important to have the latest technology, but it is necessary to take a step back and reflect on the fundamental principles that drive AI systems." Guli said. "These patterns are the engineering basis that improves the quality of the solution."
The business survival kit
For business leaders looking to stabilize their AI stack, Gulli identifies five "ripe fruit" patterns that offer the greatest immediate impact: reflection, routing, communication, guardrails, and memory. The most critical change in agent design is the step from simple "stimulus-response" bots to systems capable of Reflection. A standard LLM attempts to answer a query immediately, which often leads to hallucinations. A reflective agent, however, mimics human reasoning by creating a plan, executing it, and then critiquing its own outcome before presenting it to the user. This internal feedback loop is often the difference between a wrong answer and a correct one.
Once an agent can think, it must be efficient. This is where Routing becomes essential for cost control. Instead of sending each query to a huge and expensive "god model," a routing layer analyzes the complexity of the request. Simple tasks go to faster and cheaper models, while complex reasoning is reserved for the heavy hitters. This architecture allows companies to scale without increasing their inference budgets. “One model can act as a router for other models, or even the same model with different indications and system functions,” Gulli said.
Connecting these agents with the outside world requires Communication giving models access to tools like search, queries, and code execution. In the past, connecting an LLM to a database meant writing custom, brittle code. Gulli points out the rise of Model context protocol (MCP) as a crucial moment. MCP acts as a USB port for AI, providing a standardized way for agents to connect to data sources and tools. This standardization extends to "Agent to agent" (A2A), allowing specialized agents to collaborate on complex tasks without custom integration overhead.
However, even an intelligent and efficient agent is useless if it cannot retain information. Memory The patterns solve the "goldfish" problem, where agents forget instructions during long conversations. By structuring how an agent stores and retrieves past interactions and experiences, developers can create persistent, context-aware assistants. “The way memory is created is critical to the quality of agents,” Gulli said.
Finally, none of this matters if the agent is a liability. Railings Provide the necessary restrictions to ensure that an agent operates within security and compliance boundaries. This goes beyond a simple system message asking the model to "be kind"; It involves architectural controls and escalation policies that prevent data leakage or unauthorized actions. Gulli emphasizes that defining these "hard" the limits are "extremely important" For security, ensure that an agent trying to be helpful does not accidentally expose private data or execute irreversible commands outside its authorized scope.
Fixing reliability with transactional security
For many CIOs, hesitation to deploy agents is driven by fear. An autonomous agent that can read emails or modify files poses a significant risk if it goes off the rails. Gulli addresses this by borrowing a concept from database management: transactional security. "If an agent performs an action, we must implement checkpoints and rollbacks, just as we do for transactional security in databases." Guli said.
In this model, an agent’s actions are provisional until they are validated. If the system detects an anomaly or error, it can "revert" to a previous safe state, undoing the agent’s actions. This safety net allows companies to trust agents with write access to systems, knowing there is an undo button. Testing these systems also requires a new approach. Traditional unit tests check whether a function returns the correct value, but an agent can arrive at the correct answer through a dangerous and faulty process. Gulli advocates evaluating Agent trajectoriesmetrics that evaluate how agents behave over time.
“[Agent Trajectories] “It involves analyzing the entire sequence of decisions and tools used to reach a conclusion, ensuring that the entire process is robust, not just the final answer,” he said.
This is usually increased by the Criticism pattern, where an independent and specialized agent is tasked with judging the performance of the lead agent. This mutual verification is critical to preventing the spread of errors, essentially creating an automated peer review system for AI decisions.
Preparing for the future: from rapid engineering to context engineering
Looking ahead to 2026, the era of the single, general-purpose model is likely coming to an end. Gulli predicts a shift toward a landscape dominated by fleets of specialized agents. "I firmly believe that we will see a specialization of agents," said. "The model will still be the brain… but the agents will become true multi-agent systems with specialized tasks (agents focused on retrieval, image generation, video creation) communicating with each other."
In this future, developers’ primary skill won’t be convincing a model to work with clever writing and quick engineering. Instead, they will have to focus on contextual engineering, the discipline that focuses on designing the flow of information, managing state, and curating the context on which the model is based. "go."
It is a step from linguistic deception to systems engineering. By adopting these patterns and focusing on the "plumbing" of AI instead of just models, companies can finally close the gap between hype and bottom line. "We shouldn’t use AI just for AI’s sake," Gulli warns. "We must start with a clear definition of the business problem and how to best leverage technology to solve it."
#Agent #Design #Patterns #Missing #Link #Demos #Business
