Agentic AI Without LLMs? It Flew a Spacecraft in 1999

What is agentic AI without LLMs?

Agentic AI without LLMs is a system that senses its world, decides against a goal and acts on its own. But every goal, rule and reward in it was written by a human in advance. My definition of agentic AI says traditional IT executes instructions while agentic AI executes intent. The Five-Stage Roadmap then describes how that intent earns autonomy, from Prompted to Autonomous. So agentic AI without LLMs exists, and it has run in production for thirty years. But it only works inside a closed world that a human specified first.

New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.

You cannot interpret intent without a language model. You can decide without one.

Did agents exist before large language models?

Yes, and some of them flew. On May 17, 1999, an onboard planning agent took closed-loop control of a deep-space probe, about 65 million miles from Earth. The space agency’s report called it the first AI-based closed-loop autonomous control system to run a spacecraft. It planned its own moves and ran them. Then it found a stuck thruster and a failed electronics unit without a human in the loop.

Closer to the enterprise, a US mortgage agency’s automated underwriting system went live in June 1995. The AAAI paper that described it counted about 1,200 rules and a hard rule to decide a case in under a minute. That system has made mortgage decisions on its own ever since. Nobody called it agentic AI without LLMs then, because nobody needed the label. But it perceives a file, decides against a policy and acts.

I draw the line with a test older than any chatbot. Franklin and Graesser asked in 1996 whether a thing is an agent or just a program. Their answer: an agent sits inside a world, senses it and acts on it over time, while pursuing its own agenda. A cron job fails that test. That underwriting system passes it.

What does an agent look like without a language model?

It looks like a loop with a human-written frame around it. A rule engine matches facts to rules and fires them. Planners search from a start state to a goal state, using action models someone wrote. Reinforcement learning policies act to collect a reward someone designed. And a model predictive controller tunes a target someone set, every cycle.

The same frame shows up in every family. Humans define the state space, the action space, the goals, the rules or the reward. Then the smart part searches, learns or tunes inside that frame. That is why these systems are easy to audit and certify, and also why they break when the world drifts outside the frame.

A hyperscaler’s data-centre cooling is the cleanest live example. In August 2018 its research lab reported that its neural network had taken direct control of cooling, with operators watching. Savings rose to around 30 percent over nine months. Every five minutes it proposed actions, dropped the low-confidence ones and checked the rest against a list of safety constraints the operators wrote. Operators could exit AI control at any time. So that is agentic AI without LLMs, running a real plant, and it never read a sentence.

What does the LLM actually add?

The LLM adds the step from intent to spec, at run time. Before 2022, a human had to turn “get this request approved” into rules, states, goals and rewards. Only then could the agent run. The language model is the first technology that does that step while the system runs, so the frame can move. It reads an invoice nobody formatted, plans steps nobody coded and picks a tool nobody wired up.

But what it does not add matters just as much. It does not add exact numbers, repeatable runs or a reasoning trace an examiner can replay. Anthropic’s guide to building effective agents separates workflows from agents. In a workflow, code controls the path. In an agent, the model directs its own process. It says to find the simplest solution possible. For regulated work, that is a mandate for keeping the model away from anything that must be exact.

So my thesis holds in both directions. Instructions in, results out was IT. Intent in, outcomes out is agentic AI. Without an LLM, a human still translates intent into instructions. The engine underneath can be thirty years old and still agentic.

Where do these agents sit in the six layers?

They fill five of the six layers and skip the one that matters most today. My six-layer architecture runs intent, capability, enforcement, runtime, evidence and autonomy. The underwriting system has capability in its rules and enforcement in its policy limits. It has a runtime, an inference engine, and evidence in its reason codes. It even has an autonomy band: approve, refer, decline. What it lacks is an intent layer a machine can read. The intent lives in a human’s head and in a rulebook.

On the roadmap, the best of these systems sit at the Autonomous stage, but only inside their closed world. They earned that band the slow way. For example, the space agency ran its planner as a bounded test with a flight team on standby. The cooling lab started in advice mode in 2016, moved to watched control in 2018 and kept an operator override throughout. That is the Piloted to Governed to Assured sequence, lived out before I named it.

What breaks when there is no LLM?

What breaks is the same thing that breaks with one: the frame, not the reasoning. A home-buying platform is the case. In November 2021 the company told the SEC it was winding down the business. It had taken a 304 million dollar inventory write-down in one quarter. A pricing model that worked in a steady market bought homes above their resale value when prices moved. No circuit breaker tied purchase volume to forecast confidence.

This was a specification and guardrail failure and it involved a hallucination. In short, agentic AI without LLMs fails the way agentic AI with them fails. I keep that in mind when someone tells me the risk in agentic AI is the model. The risk was always the loop and the controls around it. That is the argument I made before the current wave arrived.

How should a regulated enterprise combine the two?

Put the LLM at the edges and fixed logic at the core, because each is good at one job. In regulated work I let the model read documents, pull out fields, settle mismatches and propose the next action. Every proposal carries a confidence score and a pointer to the source page, so a reviewer can check it. Then plain code recomputes every figure that feeds the decision before anything reaches it.

Still, the binding decision itself stays with validated models and rules. They generate the reason codes, never the LLM. That matters because an auditor will ask for the specific reason behind any decision, whatever tool made it. A reason the model wrote after the fact is a story, while a reason a rule fired is a record.

Bottom line

Agentic AI without LLMs is real, old and live in your enterprise right now. It is also stuck inside a closed world a human built. The LLM breaks that wall because it translates intent at runtime, and in return it gives up exactness. My books, Intent In, Outcomes Out and Earned Autonomy, show how I wire the two halves together, stage by stage along the roadmap.

Which of your live systems already passes the agent test, and has anyone put it on the roadmap?

Was this post useful? One click tells me what to write next.

← Back

Thank you for your response. ✨

Subscribe to my LinkedIn Newsletter

Agentic AI Architect: control design for enterprise agents.
Agentic AI Case Studies: deployment evidence, one teardown at a time.
Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
Agentic AI SDLC: the lifecycle that builds the agents.
Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.

Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.

Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.

Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.