Babak Hodjat > Quotes

“Any code or prompt written to enable an LLM to autonomously get a task done is called an agentic harness, and systems built to automatically, and even sometimes agentically, create or improve harnesses are called meta-harnesses.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“consistent. This is understandable, but defeats the whole purpose of agentification, which requires ceding some control to agents and accepting some level of inconsistency, in exchange for robustness, efficiency, and quality.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“There is no one-size-fits-all LLM for all agents. This becomes even clearer considering the advantages of using smaller, more specialized LLMs for more specialized agents. This gives us the option to run some agents on premises, for example in settings where there are sensitive data considerations. In some cases, the agent or module need not even be an LLM-based agent, and we might consider replacing it with a software module or app.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“It is important to note that the larger the number of agents, the more costly and slower the system, so there is a balance to be struck between modularity, robustness, and consistency on the one side, and cost and response time on the other.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“This also means that the bar is a bit lower on what LLM to use depending on the agent. Larger, more capable agents are expensive, slow, and mostly hosted. Being able to use smaller, more specialized LLMs means you are paying less, getting quicker responses, have the option to host agents that handle more sensitive data, and have an easier time fine-tuning the LLMs, if needed. Moreover, and as LLM context is limited, even for more powerful models boasting millions of tokens as their input context, a single agent for all uses can hit limitations. On the other hand, if context is used wisely, each agent in a multi-agent setting remains much more focused and less reliant on recalling and respecting what it is asked. The same goes for output. If you ask an LLM to produce a complex output, it is likely to fail at producing everything you request, whereas for an agent in a multi-agent setting, restricting responsibilities typically results in less output complexity requirements.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“While focused RPA implementations eventually exhausted their opportunities for marginal efficiency gains, agentic AI injects new momentum into problem-solving by introducing decision velocity—i.e., the ability to sense context, evaluate options, and act autonomously across multiple systems in near real time, without waiting for human handoffs at each step—coordinating multiple levers in a largely autonomous way. This new momentum has already translated into impressive double-digit cost savings for large organizations, as shown earlier, across their most important middle- and back-office processes.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“Agentic systems are particularly well-suited in this context and value-adding in certain types of use cases, such as: Customer support, where agents proactively resolve issues (which is not always the case with GenAI powered chatbots) instead of reactive chatbots simply escalating frustrated users to a human Supply chain optimization, where agents dynamically reroute shipments and adjust inventory levels Compliance, where agents monitor regulatory changes and continuously self-audit the organization against updated rules”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“Tasks that require reasoning, robust understanding of text or other unstructured input, or the generation of such unstructured data as output should be delegated to the LLM. More rigid rules-based behavior that requires consistency will have to be implemented in the code part of the agent.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“The gap between what generative AI promises in a presentation and what it delivers in a real enterprise workflow is rarely a model problem. It is an architecture problem, a governance problem, a trust problem.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems

“An essential tool to help apply this division of labor to a multi-agent system is sly_data, a data structure that can be operated upon by the code part of an agent, and that is passed around between agents through code, and so is not necessarily subject to LLM processing.”
― Babak Hodjat, The Agentic Enterprise: A Leader's Guide to Orchestrating, Governing, and Scaling AI Agent Systems