The Harness Is Where Symbolic AI Returned

The Standard Story About AI Agents Misses Half The System

The dominant story about building AI agents is straightforward: take a language model, prompt it well, give it tools, and let it loop. The model is doing the work. Everything else is plumbing.

This story is incomplete in a way that has practical consequences. The plumbing is not plumbing. It is a body of design decisions that determines whether the agent works in production, fails silently, or quietly degrades over time. And some of it, the parts that earn the name, is doing what symbolic AI has always tried to do: making reasoning explicit, checkable, and tied to a model of the world.

Consider what a car actually is. The engine gets all of the marketing attention, and it is genuinely impressive. But an engine on its own is just a powerful, dangerous object. The car is the chassis, the transmission, the brakes, the steering, the electronics, the cooling, the safety systems. Without them, the engine cannot become a useful vehicle. An AI agent is the same arrangement. The language model is the engine. The harness is everything that turns it into something a passenger would actually get into.

The teams we have seen building agent systems that hold up in production are, without exception, building hybrid systems. They are putting a neural reasoning engine (the language model) inside a structured layer (the harness) where the parts that bear the most weight are genuinely symbolic: typed schemas, formal grammars, structured knowledge, explicit rules. They almost never describe what they are building this way. The frame they reach for is “AI engineering” or “agentic workflows.” Neither of those names tells you what is actually being designed.

The frame that does is older than most of the people building these systems. It is called neurosymbolic AI, and it has been quietly returning through specific design choices in the modern agent harness.

What’s Actually In An AI Agent

The model is one part. The other part, the part doing more of the work than anyone admits, is the harness. The harness is where the engineering actually happens.

Each of those components has the same potential to be structured, explicit, and inspectable, the things the model itself is not. Whether they actually are turns out to be the most important engineering question in modern AI work, and it is the question the rest of this piece is about.

A 1993 Paper Got The Architecture Right

In 1993, Davis, Shrobe, and Szolovits published a paper called What Is a Knowledge Representation?. It is a classic of symbolic AI, written more than a decade before the deep learning wave that supposedly made symbolic methods obsolete.

The paper defines knowledge representation through five roles. A knowledge representation, in their formulation, is a surrogate for the real world, a set of ontological commitments, a fragmentary theory of intelligent reasoning, a medium for efficient computation, and a medium of human expression.

When a modern AI agent harness is built well, the five roles map onto its components with surprising precision.

The mapping is not an analogy. Davis and his co-authors were defining a category, and a well-built modern harness fits inside it. A typed tool schema is not like an ontological commitment; it is one. A knowledge-graph-backed memory is not like a knowledge representation; it is one.

There is an important hedge in that paragraph, the words well-built. The mapping describes the harness when its components are doing real symbolic work. Plenty of production harnesses do not satisfy the mapping at all, and the next section is about why.

What Actually Earns The Symbolic Label

The phrase “the harness is the symbolic layer” is the kind of claim that sounds good and falls apart on inspection. Plenty of components people drop into a harness are not symbolic in any classical sense, and pretending they are misleads engineers about where reliability actually comes from.

The honest position is that a harness is a mix. Some components are genuinely symbolic. Some are weakly symbolic. Some are not symbolic at all. The systems that hold up in production tend to be the ones that invest in the genuinely symbolic components at the boundaries that matter most.

Tool calling is the cleanest case in the left column, and it is worth being precise about why. A tool schema is symbolic in the same sense as a typed API: declarative, compositional, formally checkable. Classical expert systems had something extra that modern tools mostly do not, which was an inference engine doing forward or backward chaining over a rule base. Modern tool calling leaves inference to the language model. The schema is symbolic, the dispatch is symbolic, the invocation is symbolic, but the reasoning about which tool to call and with what arguments is neural. That hybrid arrangement is exactly the kind of split classical neurosymbolic AI was theorising about, even though the engineering community arrived at it by a different route.

A harness made mostly of left-column components is, by any reasonable definition, the symbolic half of a neurosymbolic system. A harness made mostly of right-column components is a clever language model application with structured plumbing. Both are useful. They behave differently in production, in ways that follow the gradient closely.

When the rest of this series talks about “the harness,” it is talking specifically about the symbolic-content version, and about the components that earn the label. The cleverer-language-model-with-plumbing version is doing something different. It tends to fail in characteristic ways, which is what the next section is about.

Three Things This Lens Lets You See

Renaming something is only useful if the new name lets you see things the old name hid. The neurosymbolic lens does, in three concrete ways that recur in client work.

The first is that it explains failure modes pure prompt engineering cannot solve. When an agent hallucinates a tool argument, the failure is not really a model failure. It is a failure of the schema, because the ontological commitment was too loose, and a failure of the verifier, because the consistency check did not catch the mismatch. When an agent forgets a critical piece of context, the failure is not just a window-size issue. It is a failure of memory representation. Teams that frame these as model problems wait for a bigger model. Teams that frame them as harness problems fix them this quarter.

The second is that it tells you where to invest. If reliability in production is the goal, the marginal next pound is more likely to pay off in the symbolic components of the harness than in the model. Better tool schemas, structured memory, formal verifiers, planning algorithms layered above the language model. These are the components classical AI spent forty years building. The literature is mature and the engineering effort is finite. The model improves on someone else’s roadmap.

The third is that it explains why scaling will not solve certain problems. A larger model improves perception, fluency, and pattern recognition. It does not, on its own, give you verifiability, composability, or auditability. Those properties come from the symbolic layer. In domains where decisions carry real consequences (clinical recommendations, financial assessments, operational safety) the verifiability comes from how the model is wrapped, not from waiting for the next release.

These patterns are not edge cases. We have seen teams stuck on the same class of failure for months because they treated it as a model problem. Investment decisions made on the assumption that the next model would close a gap that was structural. Production deployments delayed indefinitely because verifiability could not be retrofitted after the fact. In nearly every case, the missing piece was a specific symbolic component that the team had not committed to building.

The Knowledge Bottleneck Came Back, From The Other Direction

The deepest reason this frame matters is that it explains a problem the field keeps re-encountering and keeps mis-diagnosing.

In the 1980s, the expert systems movement hit what Edward Feigenbaum called the knowledge bottleneck. Systems like MYCIN and DENDRAL worked beautifully when domain experts could hand-encode rules, and they failed at the edges where rules ran out or had to be updated. The slow part of building these systems was not really writing the rules. It was being confident the rules were correct. The labour was verification, dressed up as encoding. That was the lesson the field misread, and it ended the first era of symbolic AI.

Modern AI agents have hit a structurally similar problem, from the opposite direction. The language model has eliminated the encoding labour. Prompt the model, and it produces fluent reasoning across almost any topic without anyone hand-writing a rule. But the verification labour did not go anywhere. The model produces knowledge-shaped text without anyone having committed to which version of the knowledge is correct, where it came from, or what depends on it. The bottleneck has reappeared on the output side of the pipeline.

This is why knowledge-augmented generation has become interesting again. Liang and colleagues describe an architecture that combines the language model with structured knowledge graphs, logical-form reasoning, and semantic alignment, so outputs can be traced to and verified against an explicit knowledge base. The third wave of AI, in practice, applied to the verification side of the same old bottleneck.

If a domain is one where being wrong is expensive, the lesson MYCIN taught still holds, viewed from the other side: do not put all the knowledge in the unverifiable layer. Whether that layer is a hand-coded rule base or a parameter-stored language model, the failure mode is the same. The symbolic components of a well-built harness are the response.

What Changes If You See It This Way

The most useful consequence of the lens is what it suggests for teams actually building these systems.

The first is that the symbolic investments within the harness deserve the same engineering rigour as the model, possibly more. The investments that pay off most reliably tend to be the genuinely symbolic ones. Tool schemas designed as ontologies. Memory designed as a typed representation, not as opaque retrieval. Verifiers designed to enforce formal constraints, not language-model judgement of language-model output. Planners designed to do the parts of reasoning the model is bad at.

The second is that being honest about which layer a problem lives in matters. If a problem is in the model layer, the action is to wait. If it is in the harness layer, the action is to build the missing symbolic component. Conflating the two is one of the main reasons teams ship demos and stall in production.

The third is that the symbolic AI literature is worth a fresh read. Forty years of work on knowledge representation, planning, constraint satisfaction, and formal verification is sitting on the shelf, written off as obsolete during the deep learning wave. Most of it is now directly relevant to the engineering problems modern agent teams are trying to solve. The teams that internalise this have an advantage over the ones still waiting for the model to fix everything.

The harness has been hiding in plain sight, dressed up as packaging. It is not packaging. The parts that earn the name are where symbolic AI has been quietly returning, in the components engineers reach for when they need something to hold up.

This is the first piece in a short series on neurosymbolic AI and the practical agent systems being built today. The next piece turns Davis, Shrobe, and Szolovits’ five roles into a working diagnostic for any agent system. The pieces after that examine the upgrade from retrieval augmented generation to knowledge augmented generation, and look at constrained generation as the most visible example of classical symbolic AI returning under a new name.

If you are designing AI systems for a domain where verifiability matters, a fixed-fee advisory session is built around exactly this kind of structural thinking. Get in touch to discuss what it would look like for your team.