Reliable Agentic AI

Generative AI (GenAI) started a revolution in software eating the world. It was eating the world piecemeally and rather peacefully until the end of 2022, albeit at a gradually accelerating pace. Software, as always until then, was used to automate and amplify certain areas of human activity in the same way as mechanical and electrical tools had done it before. GenAI 10x’ed that by going meta — the very way of thinking we do to plan, prioritize, and identify opportunities seemed in reach for AI. After a year or so of “prompting”, it was clear there’s a gap between the software that actually ran, and runs, the world, and the GenAI superstructure, that was mostly R&D spend on PoCs by a plethora of startups. The evolution of GenAI then took hold over the revolution, to close the gap between aspirations and reality. The reality was represented by the actual enterprise stacks, developed over decades and battle-tested in the setups where people depend on software to run reliable and help them make money or run missions critical to society (such as air traffic control) or, most commonly, both (healthcare, airfare reservations, nuclear power plant automation).

The first question that needed to be addressed is that the people who created ChatGPT were not ultimately software engineers, but former data scientists who now rebranded as AI Engineers. They play with huge datasets in pioneering ways, and their modus operandi is Exploratory Data Analysis, or EDA. It requires interactive environments supporting iterative discovery. S/R, Jupyter and Python did to data science and AI what Lisp, OCaml, Haskell and Scala did to software engineering — introducing REPL (Read-Evaluate-Print Loop), both textual and graphical, as the fundamental means of production.

But prompts will not call themselves in production. Production is not exploration. It’s making money, or deriving interest, applying the highly sophisticated, tested and optimized means of production developed in the R&D process above. The mode of this application is drastically different from the fluid scientific camaraderie of the research labs — it is mechanical scalability powered by meticulous operationalization, harnessed by a tight cocoon of observability, accountability, and operations management focused on rather primordial and brutal efficiency.

It’s almost as we have evolved a two-brain system for AI that humans had done before: the thoughtful, introspective, reflectory mind superimposed on a reptile brain acting instinctively and immediately.

Agents were the answer arrived at by the AI evolution in 2024, as a consensus answer to how AI will develop in the enterprise. In a way it is a non-answer, saying that AI will be a collection of agents, where we focus on the engineering of a distributed system while the heavy lifting of “doing AI” is now delegated to an individual agent. This is now where the prompt execution is happening, building a layer of indirection and software engineering abstraction on top of the actual “thinking” process.

On the one hand, that lets humans conveniently get busy arranging for the layout and communication of the agentic systems, which resemble the microservices that were the last software engineering wave before AI. It makes a lot of sense because intelligence is half deep thinking, half communication. And in reality the deep thinking is done by a few while the communications are done by many. Many large companies are effectively communication networks where strategy is percolated top-down from the CEO through a series of meetings to the executor nodes, which are currently highly skilled individuals. The middle managers are effectively modems, aggregating strategy from the above to suitably granular tactics for the below level. Aligning a huge organization is inherently a communication problem. The microservices faced the same dilemma, centralization vs delegation, which every company faces. The answer is often a matter of preference. The monorepo vs letting a thousand flowers bloom debate simply shows that different people will form collectives of distinct character that approach problems in differentiated ways, one suitable for centralization, the other for federation.

What’s happening now in the agentic systems is the same organizational thinking. Once an agent was conceived as a container for a prompt, or a unit of planning and thinking, it is now a highly skilled and quirky meta mind IC that requires a harness of efficient reptile managers to align and direct to doing the work desired by the business as a whole. The reptiles are the manager-supervisor agents. (I can’t help giving a shoutout to Greptile:).

Now we arrive at the problem of 2025: the agentic systems are failing. They are failing in different ways than the individual proton-agent, a prompt REPL wrapping an LLM, was failing in 2023. Because the individual REPL agents are still in flux and inherently uncertain, the communication edifice built on top of this sand of flaky agents is itself flaky.

Fortunately, the prevalence of the reptile agents points the way towards the solution. The majority of the agents in our Multi-Agent Systems (MAS) are not deep thinkers. They are the middle managers of the megacorp, keeping tabs on other agents, counting beans (packets, transitions, times, money, customers, ARR), updating databases, writing and reading logs, sending notifications, spinning up and down VMs, etc. All they really need to be are just reliable nodes in a reliable distributed system. They need to comprise Reliable Agentic AI.

I’ve noticed the pattern at the end of 2023. In 2024 at JFokus, we talked with Jonas Bonér, the creator of Akka, one of the most reliable, most developed distributed systems in production. It follows the design pattern of actor systems introduced by Erlang, the backbone of reliable telecommunications networks around the world, delivering SMS messages connecting the enormous global network of all humans and billions of devices. Akka powers such gigantic actor systems as the Oracle Cloud and the eero mesh. Jonas and I converged on the idea that an agent is inheretly an actor. An actor model defines a resource, such as a database, that the actor presides over, wraps and controls, and the only way to use the resource is by passing messages to the actor and await its responses, most often asynchronously. You’d only want to block on an actor (as in Ada rendezvous) in a really crucial situation when you cannot proceed without its answer.

The asynchronous nature of the majority of work is reflected in the Reactive architectures. Agents can do what they need by themselves, asking other agents for help when needed and keeping an eye on their response without stopping all the other jobs. This is just like us humans sending an email or a slack message and continuing with our business, inserting a result of the reply into out work in progress as needed. The problems arise when other agents fail to reply, send a wrong reply, or reply too late. These problems are similar to what we handle at work every day. If an agent is replaced, or moved from one location to another, we adjust to it. Actor systems like Akka colocate computation with data. Failed actors can be respawned elsewhere, messages rerouted automatically after failover, and multiple strategies used for backup, failover, redundancy, etc. The Reactive Manifesto encapsulated decades of experience doing all if that asynchronous orchestration with failover management reliably.

In his September 2025 AI Conference keynote, Ion Stoica focused on Reliable AI. (A longer version from the Agentic AI Summit is here.). That was a validation of our approach started in May 2025, where we convened a group of industry leaders who built reliable distributed systems for decades and now are applying their expertise to Agentic AI. The question is, how do we define reliability for AI, and which best practices of software engineering do we translate into Reliable Agentic AI definitions, and how?

Ion showed a number: about 70% of multi-agent systems fail. In a venerable Berkeley fashion, he introduced the number and a benchmark, MAST, Multi-Agent Systems Failure Taxonomy, both stating the problem and providing a research direction to work on the number. But which principles do we need to bring the number down? What foundations will be needed to build the systems bringing the number down?

This is the purpose of our Reliable Agentic AI Manifesto. Following in the tradition of the Agile and Reactive Manifestos before it, we identify the key pillars and principles of the Reliable Agentic AI systems. As its predecessors, the manifesto is short and provides a call for action to drive in the stakes, to erect the pillars, and proclaim the principles to the world — as the Molo in Venice, where the pillar of St Mark stands on the stakes driven into the mud, the lion on top holding the book of wisdom open for all to see. “Pax tibi Marce, evangelista meus,” says the open page — and the same peace will descend on the distributed systems built on the same pillars and stakes.

Read the Manifesto here. Sign it by either cutting an issue with just your name, or sending a PR to the SIGNATORIES.md file, or answering the Reliable Agentic AI Survey #1. The latter should be of interest to everyone as it lists many of the recent developments in Agentic AI architectures as choices. Write your name in the end to sign if you agree with the overall approach. We maintain the Manifesto in a reliable-ai repository on GitHub under a reasonable org — emphasizing a set of reasonable and pragmatic approaches to building AI that works for all of us.

To advance the Reliable AI ideas, we are establishing a Community Research Center for Reliable AI (CR4AI) at Northeastern Oakland, a West Coast campus of Northeastern University. As an Adjunct Research Scientist, PI and the Chief Scientist of CR4AI, I’ll be a convener and catalyst of the academic community, OSS AI engineers, startups, enterprises, and the public sector to understand and build Reliable AI for all. Two crucial areas include

• Embodied Agentic AI — where mistakes cannot be sent to /dev/null
• AI for Engineering — changing all of engineering with AI engineering

We’ll also bring Open-Source Science community to accelerate research and development with AI, reliably.

We take an engineering approach to reliability. If you have ideas on ethical, responsible, ad human-centric AI, express them in code and let us build benchmarks that will measure it. At a recent Dartmouth AI conference, I was on a panel moderated by my original academic advisor, Prof. George Cybenko, who discovered the foundational deep learning theorem that any reasonably useful function can be approximated by a neural network. Prof. Hany Farid, the pioneer of judging whether images and videos are real, was also a panelist. It’s all about Reliable AI in the end. Another speaker, General David Barnes, the former Chief AI Ethics Officer of the US Army, said that ethics is an element of reliability. We hope that engineering approach to Reliable AI will become a community-centric focus force to build and improve Reliable Agentic AI that works for all.