What is a knowledge factory?

Today, professional knowledge work (finance / economics / accounting / law) often operates much like pre-industrial craft production before the Industrial Revolution. Individual expert humans create largely manual, bespoke reports, analysis, and outputs, operating within a loose network of social contacts and relationships. [1]

Professional knowledge work still looks closer to craft production than to a standardised industrial process.

In the Industrial Revolution, one of the primary causes of the shift from craft production to industrial manufacturing was the development of machines that could increasingly replace human and animal muscle in production. Work could be broken down, standardised, concentrated, and organised around the capabilities of the machine rather than the limits of the individual craftsman.

Today, the advent of AI represents a similar kind of shift, where we have created machines (AI) that can start to replace human cognition within knowledge work. As a result, we see the same drivers that led humans to split up and standardise physical manufacturing are now leading us to similarly divide up and standardise knowledge work.

Therefore, we believe the future of knowledge work will not be defined by the current focus on "how do we make an AI drop-in replacement for a current human knowledge worker?" It will be defined by "how do we re-imagine the production process for knowledge work itself?" [2][3][4][5][6]

This is the essence of the knowledge factory Sonto is trying to build. It is about developing new processes for this new age of knowledge work. [5][6]

Right now, in the rapidly changing world of deploying AI agents, the ecosystem is concentrated around the agent providers (Anthropic / OpenAI / Google) and the harness providers (i.e. the "traditional" software and tooling provided to the agent). Increasingly, companies are trying to vertically integrate both (Anthropic -> Claude Code and Cursor -> Composer). [7][8][9]

Roughly speaking, agent providers are primarily focused on delivering increasing raw general intelligence. Harness providers are focused on providing the optimal set of tools, infrastructure, workflows, and skills that enhance agents' ability to complete tasks within a given domain. [9][10]

However, these two roles currently don't provide an obvious path to a world in which 100s-1000s of agents (+ their harnesses) can collaborate in a robust way across an array of heterogeneous tasks.

As we move from a world of largely singular agents to one of collaborative agents, we believe a crucial third role will soon emerge as one of the key drivers of AI being deployed into the real world. [6][10][11]

That new role is the knowledge factory.

Building a knowledge factory is not about improving raw intelligence or enhancing the capacity of any one agent with a specialised harness. It is about developing the system that enables any one agent with its harness to effectively and robustly collaborate within a complex multi-stage knowledge production process specialised to a specific domain. [5][6][10][11]

Therefore, in the manufacturing analogy, an agent would be the worker on the production line, the harness would be the set of tools and machines sitting at its own workstation on the line, and the knowledge factory would be the set of steps that define every specific task through the production process from inputs to outputs. [20][21][22]

Today, much of the frontier conversation is focused on ever more autonomous agents that can complete ever longer tasks, and possibly spin up agent swarms along the way. This is reflected in the current focus on the METR time-horizon evaluations as a mark of frontier progress. METR's recent work suggests frontier time horizons have been doubling very quickly, on the order of months rather than years. [12][13]

In the short term, it seems highly likely that long-running autonomous agents will soon be able to produce very high-quality bespoke work and reports (and in some cases already can). However, at Sonto we believe this is only one part of the puzzle. We believe the bigger issue is that to truly scale knowledge work AI agents will face the same coordination problem that humans have faced in physical manufacturing for hundreds of years. [5][6][18][21]

This is because long-running autonomous AI agents still seem likely to have many of the same constraints as human workers in manufacturing. [11][14][22]

• AIs also still have limited input context, and so, like humans, they still focus on one sub-task at a time. [14]
• AIs also still make mistakes, and so, like humans, quality assurance and verification steps remain key; high reliability is fundamental to successful long-run collaboration. [15]
• AIs also still benefit from specialised tooling, and so, like humans, the more constrained and well-defined the task, the more custom tooling you can bring to bear. [16]
• AIs still have not solved the search problem, and so, like humans, when faced with web-scale data, they can only sift a tiny fraction of it. [17]

For as long as these constraints remain, there will be returns to specialisation and coordination. Hence, we believe truly scaling knowledge work does not look like swarms of AI agents making infinitely more one-off pieces of work that have to largely start from scratch each time. We think it involves a production system that allows agents to perpetually build on top of, update, and combine outputs in arbitrarily long chains of work over time. [5][6][18][19][21]

Much like in manufacturing, to achieve this composability you need incredibly high-quality, standardised inputs and outputs at every step of the process. The difference between "I can cut some wood and put together a table" and "I have a largely automated IKEA production line" is a fundamentally different level of complexity. As with defects in manufacturing, when trying to scale knowledge work, as soon as bad assumptions, unspecified data definitions, extraction errors, hallucinations, varying analytical approaches, and similar defects are allowed to creep in, the system will inevitably start to collapse. [15][19][21][22]

The jump from a manual workshop to a real production line is a change in system design, not just a better tool.

To solve this coordination problem, humans since the Industrial Revolution have repeatedly driven productivity by reducing ad hoc individual human operation and judgement, and by standardising inputs and outputs as much as possible. [2][3][5][20][21][22]

It would be strange if we invented AGI (often defined as AI of human-level intelligence and capability) and the way to deploy it effectively to scale knowledge work production was to do the exact opposite of what we did with humans in every other industrial system. [2][3][5][20]

Hence, from this perspective, one of the biggest mistakes people in AI make today is to look at the current, largely artisanal state of knowledge work and think "this looks quite easy to partly automate", instead of looking at the incredible difficulty of industrialising anything and thinking "this looks a lot like manufacturing production hell". [20][21][22]

It is like walking into a 17th century home spinning setup and deciding that we will be nearly done if we just add a steam-powered spinning wheel into the person's home, instead of seeing the incredible complexity and scale that modern textile manufacturing would become - which still sees innovation to this day. [20]

Industrialisation came from reorganising the whole production system, not from dropping one stronger machine into an artisanal workflow.

If you agree with this view of the world, then we are at the beginning of a structural shift in how we perform knowledge work.

At Sonto, we are building this industrialised future. We are starting with public market equity research, a domain with abundant structured documents, clear outputs, and large gaps in coverage. We believe, though, that the underlying pattern generalises much further.

Our future blogs will dive into the details of how we believe knowledge factories will work and the specific equity research factory we are constructing first.

1. Laura Empson, Daniel Muzio, Joe Broschak, and Bob Hinings, "What is a Professional Service Firm?" (excerpt from Researching Professional Service Firms / The Oxford Handbook of Professional Service Firms), esp. p. 2 on the "bespoke nature of professional work" and intensive client interaction. Source
2. Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations, Book I, chapters 1-3 (Liberty Fund / Online Library of Liberty edition). Source
3. Adam Smith Works, "Chapter I. Of the Division of Labour." Source
4. Ronald H. Coase, "The Nature of the Firm" (1937), Economica 4(16): 386-405. Source
5. Gary S. Becker and Kevin M. Murphy, "The Division of Labor, Coordination Costs, and Knowledge" (1994), in Human Capital, pp. 299-322. Source
6. Luis Garicano and Thomas N. Hubbard, "Hierarchies, Specialization, and the Utilization of Knowledge: Theory and Evidence from the Legal Services Industry" (NBER Working Paper 10432, 2004). Source
7. Anthropic, "Claude Code." Source
8. Cursor, "Composer 2." Source
9. Google AI for Developers, "Agents Overview." Source
10. OpenAI, "A practical guide to building agents." Source
11. Anthropic, "Effective harnesses for long-running agents." Source
12. METR, "Task-Completion Time Horizons." Source
13. METR, "Time Horizon 1.1" (2026). Source
14. Anthropic, "Effective context engineering for AI agents." Source
15. OpenAI, "Why language models hallucinate." Source
16. Anthropic, "Writing effective tools for agents." Source
17. OpenAI, "BrowseComp." Source
18. Anthropic, "Long-running Claude for scientific computing." Source
19. Anthropic, "Harness design for long-running application development." Source
20. Robert C. Allen, "The British Industrial Revolution in Global Perspective" (Keynes Lecture in Economics, British Academy Proceedings 167), esp. pp. 199-200 and 212-214 on the Industrial Revolution, productivity growth, and the move from spinning wheels to mechanised textile production. Source
21. Wouter Dessein and Tano Santos, "The Demand for Coordination" (NBER Working Paper 10056, 2003), esp. abstract and pp. 1-2 on rules, task guidelines, communication, and the coordination-specialisation trade-off. Source
22. Armen A. Alchian and Harold Demsetz, "Production, Information Costs, and Economic Organization" (1972), American Economic Review 62(5): 777-795, esp. p. 777 on cooperative specialisation and the demand for economic organisation. Source