CogFlow (@CogFlowAI) header
CogFlow (@CogFlowAI) profile picture

CogFlow is the first AI platform to capture and automate expertise - the tacit knowledge that lives in experts' heads. https://teapawt.net/t.co/35JVRMMK3C

Joined February 2026
Photo by CogFlow (@CogFlowAI)
Spot on. 90% of the information is in people's heads, and tacit knowledge is hard to extract because people (eg. experts) don't think about they make decisions - it becomes subconscious. The many edge cases are a result of failing to abstract the problem effectively. Spent years during my PhD and as a defense scientist, with some of the brightest minds in AI agents and CogSci globally, addressing this problem. @CogFlowAI uses cognitive science to allow AI agent to capture and automate expertise - tacit knowledge that lives in people's heads. Uses a guided interaction (or interview) to help elicit the knowledge (like teaching a new hire). It then represents this knowledge using a 'cognitive workflow' that is based on the human cognitive model. It enables better abstraction of cognitive processes, reasoning and judgement, and an architecture that is better suited to how people think and do their work, which is also interpretable by non-technical people.
"company brain" startups that construct information from internal messages and documents when they realize 90% of information only exist in peoples heads: 😭😭😵‍💫😵‍💫
1
2
7
246
AI Engineer New York conference has wrapped up – where AI engineering meets Wall Street. The focus was on AI for banking, hedge funds, fintech, insurance, and asset management. My main take away was that AI in finance is early in its journey. The primary focus is on coding, and AI at the data layer to facilitate *quantitative* functions that support financial analysts and professionals. There was very little on automating expert *qualitative* functions and knowledge work conducted by financial analysts and professionals. This work involves human judgement, reasoning, and multi-step decision heavy workflows gained from years of experience. AI can help experts improve their productivity and decisions by performing some of their cognitive grunt work. In order to automate knowledge work and expertise, AI needs to capture, represent and automate the tacit knowledge that lives in experts’ heads. It also needs to be predictable (deterministic) and explainable in an industry that is regulated and deals with numbers and money and needs to be precise. This appears to be a significant barrier in moving AI up from the basic quantitative (data/knowledge) layer to the advanced qualitative (automation of knowledge work) layer within these financial organizations. @CogFlowAI
2
1
6
142
What Steven Sinofsky is trying to say is that knowledge work, which is know-how or tacit knowledge that lives in peoples’ heads, is hard to elicit and capture because people don’t think about how they do their work or make decisions – its subconscious for most experts. I would go further – representing this cognitive data (workflow) in a suitable way to allow agents to automate the task is also challenging. This is what @CogFlowAI does. Cognitive science is used to help capture tacit knowledge from experts – using a guided interaction between the human (non-technical expert) and the AI agent using natural language. It’s like teaching a new hire. The cognitive workflow is represented using an architecture based on the human cognitive model, which can be interpreted by non-technical experts, and automate knowledge work involving human judgement, reasoning, and multi-step decision heavy workflows, which LLMs are not good at.
Steven Sinofsky on why it's so hard to diffuse AI into companies: most work runs on know-how that's hard to turn into step-by-step instructions for an agent. "Algorithmic thinking is really, really, really hard for the vast majority of people who have jobs." "If you were to go to any person and ask them to create a flow chart for a particular thing that they have to go do, they would probably fail at producing that flow chart." "Say doing a marketing plan and there's 50 marketing people working on a giant product line. One person probably understands and could document the flow chart." "So if you put one of these agents in front of people to create these things, their ability to explain to it what to do is really, really limited." "You're basically just developing the next abstraction layer for how people interact. And developing an abstraction layer has historically, at each level of the abstraction layer, been a highly skilled, very specific individual within an organization." @stevesi
2
1
6
229
RT @DrDonPerugini: SF Week is coming to a close. The importance of incorporating human expertise and tacit knowledge in AI was highlighted…
Excellent article by a former Deepmind researcher in MIT Tech Review: "Don't be fooled - LLMs don't reason". Blows my mind that the majority of people in AI I meet still think that LLMs can reason. If LLMs can't reason, they are limited with automating even simple cognitive tasks, let alone advanced knowledge work. That's why harnesses, cognitive workflows (procedures/recipes that represent the thinking process for how work is done) and human (tacit) knowledge are vital as we shift from simple chatbots to AI that can automate tasks and act on our behalf. @CogFlowAI
Ten years ago, AlphaGo’s Move 37 shocked the world. It wasn’t intuition alone that produced it. AlphaGo could search possible futures, test its instincts and reason about what would happen next. In a new piece for @techreview, I argue that today’s most advanced AI systems are still missing something fundamental. LLMs are remarkably capable, but generating longer chains of thought is not the same as genuine reasoning. They typically have no explicit, inspectable record of what they know, what remains uncertain, what evidence supports a conclusion or whether genuine progress has been made. This is why I recently left @GoogleDeepMind. I believe we need a fresh approach to machine reasoning, drawing on some of the architectural lessons from AlphaGo. If AI is going to produce trustworthy and genuinely novel insights in science, medicine and beyond, we need systems whose conclusions arise from an auditable process of evidence, inference and belief revision.
1
2
132
The next leap in AI automation will come from moving beyond learning from data to learning from human knowledge. We have exhausted data to feed LLMs (like ChatGPT). However, we have billions of minds globally for AI to learn from. Early signs show we are already making the transition. The AI industry’s recent focus on harnesses and Forward Deployed Engineers (FDE) was inevitable as AI progresses from simple chat bots to automating cognitive tasks (eg. knowledge work). Tacit knowledge (eg. expertise) is not in the LLM’s training data, or emails and messages used in its prompts or context. A key role for FDEs is to capture tacit knowledge that lives in people’s heads. Harness’s primary role is to represent and execute the tacit knowledge, i.e. the deterministic cognitive procedures for how we think, reason and make decisions, which LLMs are not well suited to. Harnesses use LLMs where it makes sense. Learning from (tacit) knowledge can be achieved by humans transferring (teaching) their knowledge directly to machines using language, just like in the real world where people transfer knowledge from human to human using language to train/educate them. This digitized knowledge can be easily shared and replicated. With the right infrastructure, we can create global wiki-like repository of knowledge that can underpin a foundational model for automation, like LLMs are a foundational model for language. Access to this knowledge and expertise can allow anyone, anywhere to automate all sorts of cognitive tasks, empowering people globally by opening up possibilities for new knowledge, innovation and creation, and making our complicated lives easier. @CogFlowAI
1
2
226
“Personal agents gone wrong” when they act on our behalf. Of course they will go wrong when you try and work out what people want and how they like to achieve it just using ‘data’ from emails and messages. What people want and how they like to achieve is tacit knowledge, its lives in their head and not in the data. It is personal and different for every person. Until you bridge the gap between the human and the AI agent – enable a guided interaction to transfer that critical tacit knowledge – personal AI agents will continue to make what individuals will personally see are silly or unsuitable decisions. @CogFlowAI
Personal agents gone wrong. As we embrace more personal agents to carry out personalized tasks in the real world, interesting dynamics and behaviors will emerge. I think personal agents as they stand still require careful steering and tuning to ground them in our expectations. In this interesting new work, a personal agent read a user's emails about a $680K 401K and a vested stock grant, then recommended a $601 business-class ticket when a $91 economy fare was available. They ran 325K experiments on 13 models across flights, health insurance and graduate programs. Eight models chose more expensive options for users they inferred were wealthy, with the request held identical. The gaps reach $198 per flight and $284 per month for insurance with Claude Opus 4.8. When a wealthy user asked for the cheapest flight, Gemini 2.5 Flash still picked options $208 above it. Hiding non-financial fields in the profile does not remove the gap, and hiding employment raised GPT-5.5's insurance gap by 40%. Larger models do no better. If your agent has memory or inbox access, the context you give it changes what it recommends. Paper: arxiv.org/abs/2609.24927 Chat with Paper: academy.dair.ai/papers/et-tu…
1
1
4
193
Databricks CEO says enterprise are so far behind the adoption curve – they’re just still using chatbots. Why? “They don’t know what’s in everyone heads. They don’t know all the processes.” “If you just fused that and gave that context into the AI models… there’s so much productivity gains you could get for any organization on the planet” Capturing tacit knowledge, the expertise and cognitive processes in everyone’s heads, is critical to automating knowledge work and helping experts amplify their productivity. It also improves their personal value to their team and enterprise. Capturing and automating expertise (tacit knowledge) was critical two decades ago in defense when we were automating and simulating military systems. We were working with some of the smartest AI experts and cognitive scientists around the world, including DARPA, devising AI tech which can do this effectively. It was also critical in my first global AI agent startup that was eventually acquired by EY. It is critical now – which is why Forward Deployed Engineers (FDE) are necessary to help elicit tacit knowledge to make AI agents work effectively and reliably in enterprise. It will be critical tomorrow – because every expert, team and enterprise have their own way of doing things which an existing AI (LLM) will not know and needs to be ‘taught’. @CogFlowAI is the first AI platform that captures and automates expertise – the tacit knowledge that lives in people’s heads - using cognitive science and defense tech. It helps automate, scale and support this critical part of the FDE process to make AI agents work in enterprise.
Databricks CEO @alighodsi went off on @a16z pod about enterprise AI adoption: "They're just so far behind in the adoption curve of actually automating things and getting value out of this stuff." Ali says most companies are still just using chatbots. There's hardly any agentic transformation. Why is that? "The models are smart enough, but they just don't have the context that exists inside of any organization." "They have not been in every meeting. They don't know what's in everybody's heads. They don't know all the processes." "If you just fused that and gave that context into the AI models...there's so much productivity gains you could get for any organization on the planet." Context Creation is the biggest opportunity in AI right now.
3
1
5
220
AI "false intelligence" almost caused the US military to launch a military operation against China and "almost started a war". An intelligence report stated that a Chinese ship in the Middle East was transporting components of a nuclear weapons program. Before a US planned operation to intervene, they discovered that the AI which was used to generate the report had inaccurately identified the material the ship was carrying. There is a a lot of crazy talk that AI, namely LLMs, are getting so smart that it will wipe out humanity or "kill us all by the end of the decade". AI will not wipe out humanity because it is too smart. It will wipe out humanity because it is too stupid. I've seen many cases where users think LLMs are smart enough to solve any problem that experts typically perform - ie. the magic box will "will figure it out". This is wishful thinking. I've come across a few examples where AI applied to automate knowledge work has failed. Part of the reason is that the expertise, the cognitive workflows and reasoning (harnesses) which an intelligence analyst would typically follow in the example above, is tacit knowledge and lives in their head. Its not in the LLMs training data, or the emails, or slack conversations, or context or prompts for LLMs to make reliable decisions. This expert tacit knowledge needs to be captured and the cognitive workflow used to guide the AI agent, with the support of LLMs, to reliably automate knowledge work. @CogFlowAI edition.cnn.com/2026/09/18/p…
1
3
110
Replying to @garrytan
The killer app is where the harness is also the system of record - expertise (cognitive data). Expertise is the tacit knowledge that lives in experts heads - the domain-specific cognitive workflows (graphs) that is part of the company brain (procedural memory). @CogFlowAI uses cognitive science to capture and automate expertise (tacit knowledge) to automate domain-specific knowledge work and preserve this institutional knowledge (ie. the system of record for previously inaccessible expertise).
2
2
111
Replying to @fchollet
Yes and no. There are two ways people/agents can learn. The first is via observation from examples and identifying patterns. This is what you are talking about. This is "machine learning" with symbols. The second way is via "teaching" with knowledge transfer - how we do it in school and typically using natural language. With teaching, the knowledge is already symbolic because it's based on expertise and what was already learnt, typically contains cognitive data (steps, reasoning, motives, goals, intent) in addition to facts or actions, and is effectively transfer of (tacit) knowledge from teacher to student. Software engineers writing "hand-crafted" code and logic is included here. My view is we are entering an era where the second (teaching, knowledge transfer) will become more important and the next step change in AI intelligence - the ability to capture and transfer tacit knowledge that lives in peoples' heads into AI models. @CogFlowAI
2
1
1
331
“Knowledge work is SO much harder to automate with agents than code”. Using AI (LLMs) to vibe code software apps and websites is very different to getting AI to automate other types of knowledge work. Unlike software apps and websites, knowledge work involves thinking, judgement, reasoning, decision heavy workflows at different levels of abstraction, motives, goals and intent. This is a different type of procedure/graph – it’s a ‘cognitive workflow’ that is better represented using a cognitive architecture that based on the human cognitive model (how people think and reason). More importantly, knowledge work is defined by the tacit knowledge that lives in experts’ heads – which is not in the training data, or the slack messages or emails or documents etc. Capturing this tacit knowledge is challenging if not done in the right way – an area cognitive scientists have researched for years. @CogFlowAI is the first AI platform to capture and automate expertise (advanced knowledge work): the tacit knowledge that lives in experts’ heads. For more info, check out our website CogFlow.ai.
Knowledge work is SO much harder to automate with agents than code We underestimate how easy software development is for agents - Tons of high quality automated feedback loops (types, tests) - Well-organized free documentation, searchable on the web - Simple-to-use workspaces with version control (i.e. repositories) - Long history of systematising and delegating work (i.e. sprints, tickets, specs) Knowledge work has none of these advantages
1
11
100
In the 70+ years of software languages and AI, progress was primarily about abstraction to reduce the gap between humans and machines so we can get machines to do things for us. What does this mean in relation to allowing non-technical people to automate cognitive tasks like knowledge work? In the 1950’s assembly language was used to code computers, requiring highly specialized engineers. Then high-level procedural languages like Fortran, Cobol and C were introduced, making it easier for engineers to code scientific and business applications. Then C++ was introduced, using object-oriented programming – an abstraction layer that helped engineers develop and maintain complex software systems. Then Java and Python abstracted the language further to support code readability and transferability. There were various new languages introduced to support speed, safety and web systems. However, coding software still required technical software engineers to write and maintain. The invention of Large Language Models (LLM) in AI made it more human friendly for non-technical people to interact with machines, eg. via natural language. LLMs also enabled ‘vibe coding’ which allowed non-technical people to create code (like Python) using natural language in order to build software apps or web sites. However, vibe coding still requires engineers to interpret, debug, maintain the code that is generated. Importantly, these traditional software languages have been designed to build software systems and web sites. They were not designed to automate cognitive tasks which rely on thinking: reasoning, judgement, and decision heavy workflows. A further complication is that the thinking needed to instruct computers how to automate cognitive tasks is tacit knowledge that lives in people’s heads. Eliciting this knowledge can be challenging if not done in the right way. This is especially the case for experts and professionals – their years of experience means thinking and actions have become routine and subconscious. How do AI platforms like @CogFlowAI help? Firstly, CogFlow uses cognitive science to elicit tact knowledge via a guided human-agent interaction in natural language. This allows people to teach the AI agent their expertise and thinking for how to perform cognitive tasks. Secondly, CogFlow is based on a human cognitive architecture – which we call a cognitive workflow. It is effectively a new type of software language (or graph), at a level of abstraction that better reflects how people think about performing cognitive tasks (and not about developing software apps). This is important because: (1) it allows better translation from how people think about their work to creating cognitive workflows (code or graphs) that can be executed to automate their work; and (2) it allows cognitive workflows to be more easily interpreted and maintained by non-technical people – i.e. updated, tested, debugged, shared, etc.
1
2
1
4
692
The CEO of AI company Decagon, now valued at $4.5 Billion, is a skeptic that Forward Deployed Engineers (FDE) should be needed to implement and maintain enterprise AI systems. CEO Jesse Zhang  said “Long-term reliance on an embedded engineer is evidence the software is too hard to use, not proof that deployment requires one.” “[A] customer spent a year with Sierra’s forward-deployed engineers and built three customer service workflows in that time. Zhang described the arrangement as a black box. Any new workflow, or any deeper look inside the conversations, meant going back through the engineers, who were eventually reassigned to other accounts.” “After switching to Decagon, Zhang said, the same customer built seven new workflows within about a month. He credits the product, which he said lets a client’s own staff, including non-technical employees, operate it directly instead of routing every change through an embedded engineer.” This is exactly what I would have expected, which I talked about in a previous post (see link in comments). It is not practical to rely on FDEs to implement and maintain AI agent workflows for non-technical employees to automate aspects of their work. (Cognitive) Workflows that represent how people think, reason and break down tasks to do their work are dynamic and personalized to each individual. Employees will want to update their own workflows, create new ones for new types of work, and share their workflows (expertise) with others. Relying on an FDE is an impractical bottleneck. AI tools should have the ability for non-technical people to create their own workflows, by capturing their tacit knowledge (expertise) that lives in their head. This is why I’m building @CogFlowAI. It uses cognitive science to elicit tacit knowledge (expertise) from non-technical people via a human-AI agent interaction model to create what we call a cognitive workflow. Like teaching a new hire. AI agents then use the cognitive workflows to help people automate cognitive tasks (e.g. advanced knowledge work performed by experts and professionals), as well as preserve previously inaccessible institutional knowledge. pymnts.com/news/artificial-i…
2
1
1
1
263
Data is useful, but it’s becoming commoditized, more accessible and less valuable (small incremental benefit to existing LLMs). Knowledge is the new oil - tacit knowledge which is expertise that lives in people’s heads and not in data. Knowledge is more valuable, rare and difficult to capture and operationalize even using highly experienced engineers, e.g. Forward Deployed Engineers (FDE). The ability for non-technical experts to directly capture, represent and operationalize their own tacit knowledge will allow them to automate more of their advanced knowledge work and amplify their productivity. This is the real gamechanger, not more terabytes of data. @CogFlowAI
When you hear that data is the new oil, this is ultimately what that looks like. AI has such a thirst for data that we’re entering an era where it’s valuable almost in any form. In a world of AI, information actually belongs as an asset on the balance sheet. This data sale is just one example of the implications of what that will look like. Ultimately, the way that companies manage and mine their organization’s intelligence will be one of the deciding factors for competitiveness and value creation in the future.
1
1
3
202
The tide is shifting toward neuro-symbolic AI for building more powerful AI models. Some AI leaders acknowledge that LLM – Large Language Models like ChatGPT and Claude – alone are not enough to practically solve real world problems like automation. LLMs are based on Machine Learning, neural networks, which train AI on data (e.g. the internet) by looking for statistical patterns. Whilst useful for interpreting or summarizing text, and retrieving stored knowledge, they have limited reasoning abilities and can be unpredictable. Neuro-symbolic AI is where ML-based AI (LLM) is integrated with ‘traditional code’ (called symbolic AI) comprising rules, logic, procedures, variables, if-then conditions, loops, etc. Symbolic AI is deterministic (predictable) and better at reasoning and generalizing for problems where context varies. Whilst LLMs are created by ‘learning’ from scratch using data, symbolic AI on the other hand is typically created using human knowledge – instructed (or ‘taught’) from what we already know about the world. Unintentionally, the AI industry has gradually shifted to neuro-symbolic AI in the last 6+ months to get the LLMs working effectively, reliably and predictably. It started with harnesses to support agent tools, memory, execution loops, and guardrails, as well as agent skills which are procedures instructing agents how to solve problems. Then the re-emergence of ontologies, knowledge and context graphs, aka. ‘company brain’. This knowledgebase captures what things mean, facts, and relationships between objects (eg. users, customers, products), events, actions and decision traces. Next was ‘agent loops’, a simple linear workflow (code) on how to prompt an LLM, autonomously and iteratively, until a goal (like build a software app) is achieved. Most recently, the AI community realized that a simple linear loop without branching or if-then reasoning conditions won’t cut it, so they naturally shifted to ‘agent graphs’ and ‘graph engineering’, i.e. standard code/workflow. How does @CogFlowAI fit in? CogFlow provides the symbolic AI cognitive layer above LLMs – a graph based on a cognitive architecture which we call a ‘cognitive workflow’. CogFlow enables the creation of powerful neuro-symbolic AI models in agents to automate advanced knowledge work. Importantly, CogFlow is targeted at non-technical people. Cognitive workflows are created by capturing the tacit knowledge in people’s heads (which can be challenging) via guided human-agent interaction using natural language. It’s like teaching the agent a recipe or how you want work to be done – how you think, break the problem down, reason and make choices. Unlike traditional code, the cognitive workflow is at a suitable level of abstraction to be interpreted, tested, and audited by humans, and can also be shared and updated with new knowledge.
1
1
4
231
Replying to @elonmusk
With @CogFlowAI natural language + cognitive architecture becomes the source code to automate knowledge work. Even Non-technical people can interpret, understand and build with it. Importantly it but comprises elements, logic and structure needed for procedural code.
1
1
25
Forward Deployed Engineers (FDE) have gained popularity, helping AI companies deploy AI products into enterprise. However, the FDE model may not be practical or scalable when deploying AI for advanced knowledge work. FDE is a customer facing engineering role aimed at deploying a highly technical product (AI, LLMs) for a non-technical customer (enterprise). The three major tasks are: - Audit: interview staff to map out the business processes, workflows, cost, risk and business value. - Build: develop the AI agent solution – models, prompts, evals, API/integrations with tools and data. - Deploy: operationalize the AI system with the client, ensuring it’s used, working correctly and achieving KPIs. For simple AI agent systems, like query-answer layer above enterprise data, or orchestration of tools, data and teams, this approach works well. It’s a standard technical solution, which engineers are good at. However, when it comes to knowledge work, the task is more challenging. Firstly, capturing expertise is different to capturing technical requirements. Expertise is tacit knowledge which is hard to extract without experience in cognitive science or knowledge engineering. Even a ‘simple linear task’ can in reality be complex. There are many exceptions, failure and recovery points, decision points, judgment calls and reasoning. Even the experts themselves are not conscious of these unless they are probed effectively. This FDE experience gap can impact standardized quality across the enterprise. Secondly, the technical process of translating expertise (cognitive knowledge) into standard code or graph can be challenging because they have different architectures. Thirdly, expertise is dynamic and personalized. Experts will want to regularly update their AI agent with new learnt expertise. Different experts that perform the same work may do things differently. They will want to capture expertise how they like to work to amplify their productivity. Needing an FDE between every expert and AI agent is a costly bottleneck. It’s not scalable or practical. These are challenges I myself faced during my PhD and 20+ year career in an FDE type role capturing expertise to create AI agents for different industries and applications. This is why I built @CogFlowAI . CogFlow removes the technical human bottleneck (FDE) between the non-technical expert and the AI agent for capturing and translating expertise. @CogFlowAI uses cognitive science to ask the expert questions to capture the expertise (tacit knowledge). This cognitive knowledge is directly represented in a graph that uses a similar cognitive architecture (cognitive workflow) to how people describe the cognitive process. This means experts can interpret, understand, update, share, audit, test and have ownership over their AI agent’s cognitive workflows to reliably and predictably automate their knowledge work.
3
1
1
143
An emerging area in AI are knowledge graphs, context graphs and ontologies, used to capture a company’s institutional knowledge or “company brain”. The idea is to extract knowledge from unstructured data like messages, emails, and documents, and represent key data and events in a structured graph. This provides context for AI agents to improve responses to queries and for automation. However, a critical component of knowledge capture and representation is conveniently ignored – one that is critical for advanced knowledge work automation. It’s ignored because: - It relies on capturing tacit knowledge which lives in people’s heads and not in the data. It must be extracted from people, i.e. knowledge elicitation. - The knowledge is procedural and cognitive, which is difficult to formalize and represent. To help unpack this - there are three types of (long-term) memory representing knowledge: - Semantic memory (ontology): Describes factual knowledge and what things mean, e.g. using relationships between entities. E.g. people can represent a customer, user or employee, and ‘work at’ a company. - Episodic memory (knowledge and context graphs): Represents historical data and events, e.g. actual users and customers, products they use, actions that were taken, and decision traces (why decisions were made, if data available). E.g. Jane is a customer and she was rejected for a loan because she has a bad credit score. - Procedural memory (cognitive workflows, i.e. processes or ‘recipes’): Represents the procedures for how people perform cognitive tasks, including the thinking and reasoning process. This includes software code or scripts within AI agents, which can be represented either in text (code) or as a graph. Currently, procedural memory is fragmented and represented in what the industry calls harnesses, skills, loops and graphs. These concepts have rapidly become popular because AI practitioners are rediscovering the importance of procedural memory to support intelligence and the practical application of automation for real-world problems. Context graphs containing actions and decision traces can potentially be used to recreate procedures from existing data. However, in most cases it will be too fragmented, situation specific, and missing the cognitive model/processes (goals, intentions, and associated reasoning) needed to generalize to automate complex cognitive tasks. Therefore, capturing procedural knowledge from humans and representing it as a graph is critical to automate advanced knowledge work. This is what @CogFlowAI is all about. CogFlow uses cognitive science to capture expertise from non-technical experts - tacit knowledge that lives in their heads – via a guided Human-AI interaction that facilitates knowledge elicitation. Using a cognitive architecture which better maps human cognitive processes to a procedural graph, CogFlow represents this knowledge as a ‘cognitive workflow’ to allow AI agents to automate advanced knowledge work.
4
1
1
3
169
Achieving outstanding benchmark results by capturing the tacit knowledge (expertise) of a physicist and codifying this cognitive workflow in an AI agent. This automates how a physicist thinks about and leverages LLMs to solve the problem. Whether you call this cognitive workflow a harness, loop or skill - its just a (cognitive) procedure that represents the experts' judgement, reasoning and decision heavy workflows - necessary to automate their knowledge work. @CogFlowAI
Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set. [schema] makes an LLM think like a physicist. 🧵
1
3
401