The word “agent” is doing too much work in 2026. It describes Harvey’s legal document processor and a hypothetical system that autonomously manages your entire business. It describes GitHub Copilot completing a function and a theoretical AI researcher that designs its own experiments. It describes a customer service chatbot that routes tickets and a future entity that replaces your entire support team.

These are not the same product. They are not on the same capability curve. They do not share the same economics, the same risks, or the same timeline. Treating them as one category is the central confusion of the current AI market, and it is driving billions of dollars in misallocated capital.

Two Categories, One Name

Narrow task agents execute predefined workflows with known inputs and expected outputs. They follow a script, call specific APIs, process specific data formats, and produce specific deliverables. A legal document review agent reads a contract and flags non-standard clauses. A code completion agent predicts the next line based on context. A data extraction agent pulls structured fields from invoices.

These agents work. They work reliably enough to process real enterprise workloads at scale. Harvey runs 25,000 of them across 1,300 law firms. GitHub Copilot contributes to real codebases at companies that would notice if the code were wrong. Customer service agents handle real tickets for real customers.

The reason they work is that the task structure is defined by humans in advance. The agent does not need to figure out what to do. It needs to execute what it was told to do, efficiently and accurately. The boundaries are hard-coded. The failure modes are known. The human oversight is built in.

General autonomous agents would operate in open-ended environments with undefined goals, novel situations, and no predetermined workflow. They would figure out what needs to be done, decide how to do it, execute the plan, adapt when things change, and handle situations their designers never anticipated. This is what the term “agent” implies to most people reading marketing material.

These agents do not exist. Not at any price point, from any vendor, running any model.

The ARC-AGI-3 Evidence

ARC-AGI-3 is the first benchmark that directly measures the capabilities general autonomous agents would need. It drops an AI into a novel environment with no instructions, no goals, and no predefined workflow. The agent must explore, learn the rules, infer the objective, plan a strategy, and execute it. The environments are simple 64×64 grids that children solve in minutes.

Frontier LLMs scored under 1%. The best purpose-built agent scored 12.58%. Humans scored 100%.

That 100:1 gap between human and AI performance on tasks requiring genuine autonomous operation is the single most important data point in the AI agent market. It tells you exactly where the boundary lies between what agents can do (narrow tasks) and what they cannot do (general autonomy).

The benchmark also reveals why the gap exists. Current AI agents lack persistent memory across interaction steps. They cannot form and revise hypotheses based on feedback. They cannot explore efficiently. They cannot infer goals from context. These are not incremental improvements away from being solved. They represent missing capability classes that no amount of scaling (bigger models, more data, more compute) has yet addressed.

Why the Conflation Matters for Valuations

Harvey is valued at $11 billion on $190 million ARR. That is a 58x revenue multiple. The multiple prices in a future where Harvey’s agents do progressively more autonomous legal work, expanding the revenue per customer and the total addressable market.

If Harvey’s agents remain narrow task executors (excellent at contract review, document extraction, and compliance checking within predefined workflows), the business is real but bounded. At 58x revenue, the market expects Harvey to grow revenue by 5 to 10x over the next few years. That requires either many more customers (Harvey already serves most of the AmLaw 100) or much more revenue per customer (which requires agents doing more types of work autonomously).

The same logic applies across the agent economy. OpenAI’s enterprise push, Anthropic’s $19 billion ARR, and the $200+ billion invested in AI startups since 2023 all assume agent capabilities will expand from narrow task execution toward general autonomy. If that expansion happens gradually over 3 to 5 years, current valuations look reasonable. If the ARC-AGI-3 gap represents a plateau that takes a decade to close, the valuations are ahead of the technology by years.

Nobody knows which scenario is correct. But the distinction between the two is worth trillions of dollars, and most market commentary does not even acknowledge it exists.

Where Narrow Task Agents Create Real Value

The economic case for narrow task agents is strong and does not depend on any future capability breakthrough.

A legal associate billing $500 per hour spends 60% of their time on tasks that a narrow agent handles well: reading contracts, flagging deviations from standard terms, extracting key dates, cross-referencing clauses. If an agent handles that 60% at $50 per task, the law firm saves $250 per hour of associate time. At 1,300 firms, that saving scales to billions.

A software engineer spends 30-40% of their time on boilerplate: writing tests, implementing known patterns, completing functions from context. Copilot and similar tools handle this at near-zero marginal cost. The engineer focuses on architecture, design, and novel problem-solving. The productivity gain is 20-40% in the tasks where the agent helps, which translates to 10-15% overall productivity improvement.

Customer service agents that route tickets, generate first drafts of responses, and handle FAQ-level queries reduce support costs by 30-50% for companies that deploy them correctly (with human escalation for complex cases).

None of these use cases require general autonomy. They require good pattern matching, reliable execution within boundaries, and tight integration with existing systems. The value is real, measurable, and growing.

What General Autonomy Would Actually Require

Closing the ARC-AGI-3 gap requires capabilities that current architectures do not provide.

Persistent, updateable memory. An agent operating autonomously needs to remember what it has tried, what worked, and what failed, not within a single session, but across days and weeks. Current LLMs have context windows that serve as short-term memory but no mechanism for long-term learning from experience.

Active exploration. An autonomous agent in a novel environment needs a strategy for gathering information efficiently. Current models respond to prompts. They do not proactively seek information, design experiments, or allocate attention based on uncertainty.

Goal inference. Real-world tasks rarely come with clearly stated objectives. An autonomous agent must infer what the user actually wants from incomplete, ambiguous, and sometimes contradictory instructions. Current models follow instructions literally or hallucinate intent.

Self-correction under uncertainty. When an autonomous agent encounters a result it did not expect, it needs to update its model of the world and adjust its strategy. Current models either continue with their original plan (ignoring the feedback) or restart from scratch (wasting the work done so far).

Research is active in all four areas. Progress is real but incremental. The Helsinki team’s graph-based agent that scored 12.58% on ARC-AGI-3 shows that purpose-built architectures with explicit memory and exploration strategies can outperform bare LLMs by an order of magnitude. But 12.58% is still a long way from 100%.

The Honest Framing

The AI agent market in 2026 is selling narrow task agents. The AI agent market in 2026 is valued like it is selling general autonomous agents. The gap between those two statements is the single largest risk in the technology sector.

This is not a criticism. Narrow task agents are a real, growing, commercially valuable product category. The companies building them are solving real problems for real customers. The revenue is genuine. The productivity gains are measurable.

But the pricing assumes a future that the technology has not yet delivered. Every investor, every enterprise buyer, and every founder in the agent economy should be asking one question: am I building (or buying, or investing in) a narrow task agent or a general autonomous agent? And if the answer is narrow, is the valuation priced for narrow?

The companies that answer this question honestly will navigate the next three years successfully. The ones that conflate the two categories will discover the distinction the hard way.