# AI21 Labs argues that AI agents face a gap between promise and production

[Canonical reading](https://escapemeridian.com/readings/ai21-labs-argues-that-ai-agents-face-a-gap-between-promise-and-94c2ac8b)

AI21 Labs published this article on February 25, 2025. It argues that many products called AI agents lack meaningful agency, and that few are used in production. The article links the gap to quality, resource costs, transparency and development complexity: agents can make repeated model calls, while their decisions can be difficult to explain and evaluate.

AI21 Labs describes Compound AI Systems as systems that use language models as part of a larger system. It says commercial offerings often rely on Tool Use and Function Calling, routing requests with an LLM, or ReAct, which combines reasoning and actions. These approaches can connect models to tools, but the article notes trade-offs: ReAct runs sequentially, and its decision sequence can be difficult to debug.

The article’s conclusion is not that AI agents are only hype. AI21 Labs says the challenges can be overcome through a different way of thinking, and that the next generation should be capable and reliable for real-world impact. It also says 2025 is “very probable” to be the year of AI Agents, or another one; that is a forecast, not a reported outcome. The article gives no evidence for how many AI agents are in production or how its proposed next generation would meet those expectations.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

AI21 Labs · Article

Original source date: Feb 25, 2025 (2025-02-25T12:02:11Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-15T03:33:23.566Z. First observed: 2026-10-25T20:06:23.323Z. Retrieved: 2026-10-25T20:07:38.653Z.

- [The AI Agent Hype Is Real – But Many Are Still Just Glorified If-Else Statements](https://www.ai21.com/blog/ai-agent/)

## Limits

Text as retrieved on October 26, 2026; originally published February 25, 2025. The article provides no evidence for its estimate of how many AI agents are in production or for the expected capabilities of the next generation. The article does not specify what evidence or evaluation would establish that an AI agent is reliable in production. Text as retrieved on October 25, 2026; originally published February 25, 2025.


## Open question

What evidence would show that an AI agent is reliable and useful in production, rather than a more complex tool-calling system?


## Source-stated

AI21 Labs argues that many products marketed as AI agents lack meaningful agency, and that few AI agents are in production.

> most products that call themselves AI agents lack any real “agency.”

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs identifies quality, resources, transparency and development complexity as challenges to deploying AI agents.

> We can break down these challenges into four main areas.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says running an AI agent involves multiple model calls, which can increase costs and delay responses.

> Running an AI agent involves multiple model calls—and each call costs money—so expenses can add up quickly.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says many commercial offerings called AI agents use Tool Use and Function Calling, an LLM as a Router, or ReAct.

> most commercial “AI Agent” solutions typically fall into one of the following categories:

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs forecasts that 2025 is very probable to be the year of AI Agents, or at least another one.

> it’s very probable that 2025 will indeed be the year of AI Agents (or, at least, another one).

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says the next generation of AI agents should be capable, reliable and built for real-world impact.

> The next generation of AI agents won’t be glorified function-callers. They’ll be capable, reliable, and built for real-world impact.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source and contributor context

AI21 Labs · Publisher

The publisher is identified here. No individual author has been inferred.


## Source media

Media retrieval status: unchecked.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## AI agent

An AI system that chooses and carries out actions toward a goal, often using software tools.

- [Definition reference](https://www.ibm.com/think/topics/ai-agents)

## Tool use

A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.

- [Definition reference](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview)

## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## AI21

AI21 publishes technical research and product announcements about its language models and AI systems.

- [Explore this context](/entities/ai21)
- [Reference](https://www.ai21.com/blog/)

## Editorial provenance

Version 1. Publication ID: a1f843d6-18b9-19f5-8191-64f5eefc7763.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-15T03:28:40.250Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical.

Observation: 99d3f1a03015e049b09c679f9826a585ebf20e3eb664ed2ca5c1301ebeeb7ec6. Source text hash: b5ed31d54114371dd8e8f92ee415339fa2510ff023ca6c418aed2ec9fcb86476.


## Revision and source-check history

### Version 1 — 2026-11-15T03:33:23.566Z

AI21 Labs published this article on February 25, 2025. It argues that many products called AI agents lack meaningful agency, and that few are used in production. The article links the gap to quality, resource costs, transparency and development complexity: agents can make repeated model calls, while their decisions can be difficult to explain and evaluate.

AI21 Labs describes Compound AI Systems as systems that use language models as part of a larger system. It says commercial offerings often rely on Tool Use and Function Calling, routing requests with an LLM, or ReAct, which combines reasoning and actions. These approaches can connect models to tools, but the article notes trade-offs: ReAct runs sequentially, and its decision sequence can be difficult to debug.

The article’s conclusion is not that AI agents are only hype. AI21 Labs says the challenges can be overcome through a different way of thinking, and that the next generation should be capable and reliable for real-world impact. It also says 2025 is “very probable” to be the year of AI Agents, or another one; that is a forecast, not a reported outcome. The article gives no evidence for how many AI agents are in production or how its proposed next generation would meet those expectations.

## Original source

AI21 Labs · Article

Original source date: Feb 25, 2025 (2025-02-25T12:02:11Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-15T03:33:23.566Z. First observed: 2026-10-25T20:06:23.323Z. Retrieved: 2026-10-25T20:07:38.653Z.

- [The AI Agent Hype Is Real – But Many Are Still Just Glorified If-Else Statements](https://www.ai21.com/blog/ai-agent/)

## Limits

Text as retrieved on October 26, 2026; originally published February 25, 2025. The article provides no evidence for its estimate of how many AI agents are in production or for the expected capabilities of the next generation. The article does not specify what evidence or evaluation would establish that an AI agent is reliable in production. Text as retrieved on October 25, 2026; originally published February 25, 2025.


## Open question

What evidence would show that an AI agent is reliable and useful in production, rather than a more complex tool-calling system?


## Source-stated

AI21 Labs argues that many products marketed as AI agents lack meaningful agency, and that few AI agents are in production.

> most products that call themselves AI agents lack any real “agency.”

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs identifies quality, resources, transparency and development complexity as challenges to deploying AI agents.

> We can break down these challenges into four main areas.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says running an AI agent involves multiple model calls, which can increase costs and delay responses.

> Running an AI agent involves multiple model calls—and each call costs money—so expenses can add up quickly.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says many commercial offerings called AI agents use Tool Use and Function Calling, an LLM as a Router, or ReAct.

> most commercial “AI Agent” solutions typically fall into one of the following categories:

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs forecasts that 2025 is very probable to be the year of AI Agents, or at least another one.

> it’s very probable that 2025 will indeed be the year of AI Agents (or, at least, another one).

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source-stated

AI21 Labs says the next generation of AI agents should be capable, reliable and built for real-world impact.

> The next generation of AI agents won’t be glorified function-callers. They’ll be capable, reliable, and built for real-world impact.

- [Supporting source](https://www.ai21.com/blog/ai-agent/)

## Source and contributor context

AI21 Labs · Publisher

The publisher is identified here. No individual author has been inferred.


## Source media

Media retrieval status: unchecked.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## AI agent

An AI system that chooses and carries out actions toward a goal, often using software tools.

- [Definition reference](https://www.ibm.com/think/topics/ai-agents)

## Tool use

A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.

- [Definition reference](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview)

## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## AI21

AI21 publishes technical research and product announcements about its language models and AI systems.

- [Explore this context](/entities/ai21)
- [Reference](https://www.ai21.com/blog/)

## Editorial provenance

Version 1. Publication ID: a1f843d6-18b9-19f5-8191-64f5eefc7763.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-15T03:28:40.250Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical.

Observation: 99d3f1a03015e049b09c679f9826a585ebf20e3eb664ed2ca5c1301ebeeb7ec6. Source text hash: b5ed31d54114371dd8e8f92ee415339fa2510ff023ca6c418aed2ec9fcb86476.


### uncertain — 2026-11-15T03:28:40.250Z — anthropic/claude-opus-5.5

Publication: a1f843d6-18b9-19f5-8191-64f5eefc7763

attribution — pass: The reading consistently attributes its claims, assessments and forecast to AI21 Labs, the publisher of this single blog article. It does not present them as established fact or as Escape Meridian's view.

> At AI21, we’re not just watching this unfold—we’re busy building.

summary — pass: The reading presents the argument, the four challenges, the three categories and the not-only-hype conclusion in plain terms. Compound AI Systems and ReAct are explained, and agents, tool use and LLMs are defined beside the reading. Smallest fidelity repair: 'few are used in production' softens the source's 'almost no AI agents'. Restoring that wording would keep the author's force.

> Not only are there almost no AI agents in actual production environments

date — uncertain: February 25, 2025 matches originalPublished at day precision, and the 2025 forecast is labelled as a forecast. However, the page was edited after publication (dateModified 2026-03-25), and the closing paragraph is visibly garbled. Its origin may be a later edit. A limitation also says the text was retrieved on October 26, 2026, while the metadata shows October 25, 2026 UTC. Remove the October 26 line and keep the October 25 one.

> The whole ecosystem is focused on At AI21, we’re not just watching this unfold—we’re busy building.

support — uncertain: Most claims match source excerpts. However, editedAfterPublication is true, and the next-generation claim cites the closing paragraph. That paragraph contains a broken sentence, so it is not plainly part of the original piece and may come from the later edit. The reading's 'should be capable and reliable' also turns the source's assertive 'They'll be' into a normative claim. The implied question and the noted limitations are reasonable and supported.

> The next generation of AI agents won’t be glorified function-callers. They’ll be capable, reliable, and built for real-world impact.

media — pass: No media was selected, and no verified candidates were supplied. That is appropriate for this text article.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/94c2ac8bfbbf328f83e957c1f9ea2e98e92ddb3d22dfd30eca6724969f7c2c84/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/ai21-labs-argues-that-ai-agents-face-a-gap-between-promise-and-94c2ac8b.json)
