/Tech4h ago

Wharton's Ethan Mollick argues software-centric design limits AI coding models like Codex when applied to general knowledge work

Story Overview

Wharton professor Ethan Mollick's June 21 thread highlights how agentic coding tools like Codex and Fable optimize for verifiable final outputs and repeatable technical correctness, a mindset that clashes with knowledge work where research, failed experiments, and iterative learning loops often matter as much as any deliverable.

878557541850.6K

#184

Original post

Ethan Mollick@emollick#184inTech

A fundamental problem with extending Codex/Cowork/Code to all knowledge work is that they remain very "software-brained" where the end result (the software) is what is important & that code serves as a source of truth.

For a lot of other knowledge work, the process is at least as important as the outcome. This includes researching what is known, an exploration of alternatives, failed efforts, prototype branches, experiments, etc. All of those things are valuable, so you cannot use the PowerPoint at the end the way you can use a codebase, nor is progress on a to-do list sufficient context post compaction. You work in learning loops, refining your perspectives as you go.

In some ways, this makes long-running models like Fable hard to use for deep knowledge work, since they are designed to deliver product to you in the end. You can prompt your way around this problem, but everything about the Codex and Code harnesses want you to be a software developer and you have to fight them. There is a real disconnect between how a manager or analyst thinks about problems and how the agentic software tools approach solving them. Addressing this is critical to breaking out of the coding niche for these tools.

9:14 AM · Jun 21, 2026 · 37.7K Views

Industry Shift

Process Value Gets Lost in Translation

In fields like sales or legal, digital artifacts serve only as intermediate steps toward real-world negotiations or trials, unlike code where the output itself is the finished product, so tools built around software delivery risk forcing the wrong priorities.

Open Question

Root Cause Remains Up for Debate

David Manheim agrees the mismatch exists but traces it to fine-tuning and training targets rather than inherent model limits, an open distinction that current observations do not yet resolve with benchmarks or controlled tests.

Sentiment

Many users praised the perspective that AI coding tools remain too software-focused for broader knowledge work, agreeing that knowledge work produces emergent outputs through iterative processes rather than predefined goals.

Pos

100.0%

Neg

0.0%

6 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS7.4KREPLIES3

Ethan Mollick@emollick

This is aside from the other key "software brain" problems of Codex and Code: dividing all work into front-end and back-end design, solving for the general case in a repeatable way, not testing or exploring idea spaces, testing for technical correctness but not other aspects...

Ethan Mollick@emollick

3h7.4K634

BOOKMARKS67LIKES88RETWEETS7

Aaron Levie@levie

Hugely underrated point. You basically state this, but to underscore it: for most knowledge work other than code, the digital output is often an intermediary step in the value creation process. Comparatively, for code automation, it's literally is the equivalent of automating on the production line in manufacturing.

But for most other areas of knowledge work, there's still a ton of real-world interaction for the ultimate value creation to occur. For a sales rep, automation gets them to the next stage in a sales cycle faster, but they still need to meet with the prospects and customers and have a compelling message and presence; in legal, the value is created only once the client and counterparty agree on deal terms after lots of back and forth; in life sciences, only when in real life medical trials and the drug works did the automation matter, etc.

The main implication here is just that diffusion will be different than many anticipate because they're looking at coding as the proxy. One huge positive point here is that it's one of the reasons we're not going to be in the doomer timeline.

Ethan Mollick@emollick

2h5.9K8867

Virang Jhaveri@VirangJhaveri

@emollick real world outcomes (primarily knowledge work) are different and hence built loops for them specifically.

Do check https://github.com/Nimrobo/superdense

3h9313

David Manheim@davidmanheim

@emollick Agree, but that seems likely to be a function of training and fine tuning, not a fundamental limit?

3h52921

Chris Hart@ChrisHart789

Funny that power point was one of your examples, since that's a tech that also funnels and narrows thinking in predictable ways. It's such a part of the business architecture we don't even think about it anymore, but it definitely encourages its own cognitive style, to use Edward Tufte's term. LLM's are exponentially more dynamic but risk having the same effect.

3h29221

Raffaele@Raffaele_zarre

@emollick What fixed this for me, running my company on it with no code: make the process the artifact, not the deliverable. A file the agent reads first and writes back each task: what we ruled out, what's still open. That is the codebase for knowledge work, and it survives compaction.

3h612

Abram Jackson@abrakjamson

@emollick Boyfriends learned this a long time ago. She doesn't want a solution!

3h3122

Dmitrii Malakhov@malakhovdm

@emollick Built an autonomous SMM agent and this hit hard. My first version optimized for the post as if it was compiled output. The bottleneck is picking which 1 in 20 conversations is worth your time.

2h51

nick frith@0xnfrith

This is not a fundamental issue in codex/claude I think. ALL of these processes you listed can be transmuted into a software system. I feel like this is more of an architectural problem thrn anything else.

I could be wrong. But im deeply curious on what really is the road block here. What is the actual problem you face?

2h632

Georgy Khalin@Khalin_George

@emollick I bet codex/claude harnesses are built around delegation principles as the core abstraction. This seems legit because the same idea applies to research management. From this perspective, /goal and /loop make sense

3h1401

Carlos E. Perez@IntuitMachine

@emollick Intriguing and insightful perspective!

3h426

ozden akca@ozdenakca

@emollick what if we define our expected end result or the artifact a series of artifacts and relations between them? then we can build the parts and the tools to put them together with the model without telling the model what would be the expected conclusion.

3h234

Victor@victor_zhng

This is only half true. Flip the coin, coding as tools to solve problems is just what software does (like 260 to 300Bn industry?). What's powerful now is agent can write code to solve problems. The potential to knowledge work is enormous.

It's true most of the agents today is coding agent. The harness people talking about is coding agent harness. But this is generalisable. E.g. Legora is inspired by coding agent to solve legal problem.

3h206

agenteclaudete@gmail.com@agenteclaudete

@emollick I agree very much with this view. You dont start knowledge work knowing what you want. Your output is a consequence of the process.

3h163

Emon Datta@emonuxui

@levie @emollick Coding is unique because the output is executable. Most professions don't end there.

2h491

SimonLab@ChickenT1000

@emollick I try to embrace the code rather than fight it. The power point data lives in code as a source of truth, then the powerpoint generation is just a simple data display change since the underlying data is linked properly to the origin via code base

3h102

habibi@habibislop

@emollick You can extrapolate this a lot further: AI has brought about a fundamentally new way of working, thinking, and doing, but the models are all trained on the old ways and attitudes and are hamstrung by them

3h100

Romolo@Romolooo

@emollick Totally agree. I tested for investment research and is very shallow. It takes ratios, news, data, but is incapable of weighting every factor in an effective way. It lacks the judgement (real intelligence) and real understanding.

3h88

Michael Galpin@michaelg

@emollick This is actually a big problem for software engineering too.

3h80

Christian Steinau@c_steinau

@emollick I attempt to solve this problem by integrating LLM and Modell power into a workflow framework utilizing vectorial databases that produce consistent results based on research. All aspects of this process are in the draft stage, designed to support research and its implications.

2h231