Agents need their own spreadsheet

Originally published at Synthesis Labs · 30 September 2026

Agents don’t need more context. They need a workplace.

For the last few months I’ve been throwing agents at pretty much every bit of knowledge work I can find - sales reporting for our exco, figuring out who attended which training sessions, hunting down unused software licences (in addition to my day-job, where I’ve committed over 4,000 contributions - github.com/drshade for proof!). And honestly, the results are often amazing. An hour with an agent and you have something that would have taken a person a week.

But then next month comes around.

You open a new session, explain the problem again (a bit differently this time, because who remembers exactly what they said), paste in the new data, and get something… different. Different groupings, different caveats, a different total. Neither version is obviously wrong - but nobody can tell you why they differ, or which one to trust.

The work was great. It just lived on a whiteboard - and at the end of the session somebody wiped it.

Who wiped the whiteboard?

Nobody, really. That’s just what a chat session is. Everything the agent knows sits in its context window, nothing it produced was checked against any kind of structure, and “run it again” actually means “ask again and hope”.

So you get clever and just keep using the same session. Month after month, same chat, same agent, all that lovely context still in there. Until one day, halfway through something important, you see the dreaded “conversation compacted” - and the agent wipes the whiteboard for you :)

Chat is a great place to think. Keeping what you’ve thought is another matter.

A new “spreadsheet”

Spreadsheets are not glamorous (sorry, finance friends), but they gave knowledge workers something we now completely take for granted. The rows and columns say what the data is. The formulas say how it’s derived. Next month you drop in new numbers and the whole thing recalculates - the same way, every time. And when you move on, the sheet stays behind for someone else to open, follow and argue with.

Here’s the bit I love. When the spreadsheet arrived in 1979, accountants never learned to program - they learned to model. The tool asked for a new skill, and that skill ended up defining the next forty years of knowledge work. I think we’re at exactly the same point again. Teams won’t learn to code for their agents. They’ll learn to name their work precisely: the kinds of things, the fields, the links and the rules their agents operate on.

My strong belief is that agents need their own spreadsheet. Not a longer context window - a workplace. Somewhere with structure, memory and a review step, where the work survives the session.

And it matters more now than ever. Investors talk about beta and alpha. Beta is the return you get just for showing up - buy the whole market and everyone gets the same. Alpha is the extra you earn from knowing something nobody else does. Agents work exactly the same way. Frontier models already know the beta - the playbooks, the public docs, the best practice that every model ships with. What they don’t know is your alpha: why the deal actually closed, which supplier always slips, what the customer really meant. That’s the edge only your business has, and it decays fast and hides in inboxes and people’s heads. An agent with nowhere to keep it can only ever do the beta.

Carve it out first

This doesn’t mean taking agents out of the loop. It means changing what the conversation is for. The most valuable hour you can spend with an agent is carving out the structure: what are the things we care about, how do they relate, what are the rules, and where does the evidence come from?

And this matters MUCH more than it sounds. While building that exco sales report I ran into a CRM rule that decided which division owned each deal. The rule was too blunt - one project spanning five business units landed entirely on a single division. No agent, however clever, could have reported that correctly, because the structure underneath was wrong. The fix was in the structure, not the prompt.

Get the structure right and the agent’s job gets much smaller. And much more reliable.

Enter Kyyn

This is exactly the problem we’re building Kyyn to solve. It’s in closed beta and we’re still testing it hard (so go easy on us!), but the idea is simple: give agents a proper place to do knowledge work.

A Kyyn knowledge base is a Git repository of typed records that carries its own schema - the rows and columns, if you like. The agent reads evidence from wherever it happens to live - mail, meetings, documents, Salesforce, finance systems, Git commits, a tidy structured store or a messy folder of dumps (it really doesn’t matter) - and proposes changes. A human reviews the actual diff, and only accepted facts make it in, each one traceable back to where it came from.

Once the structure and the process are carved out properly, they stop needing you in the chair. That sales report now rebuilds itself every day from the CRM, the billing schedule and the budgets, with the same rules every time. Nobody has to remember what prompt they typed last month.¹

Slow to build, fast to run

There’s a nice split of labour hiding in here too. Carving out the structure is slow, deliberate thinking - which is exactly what large language models are good at in conversation. But once it exists, most of what’s left are small, well-posed questions: did this person actually attend the training, or were they only invited?

Those don’t need a big model thinking out loud. They suit fast models that return a calibrated probability, with the threshold written into the process instead of buried somewhere in a prompt. Slow thinking to build the spreadsheet. Fast thinking to fill in the cells.

So here’s my question for you: how much of the work your agents did last month could you run again today and get the same answer?

If the honest answer is “ummm, not much” - I’d love to chat. Kyyn is in closed beta, and real, repeatable knowledge work is exactly what we want to throw at it.

Footnotes

1. There’s some other fun stuff brewing in Kyyn too (still very much work in progress). We’re adding support for both System One and System Two style AI models - fast judgement models like Jev that return a calibrated probability, alongside the big LLMs that reason slowly and deliberately. And we’re building agentic flows as type-safe, composable blocks, so you can see a whole flow before it ever runs. That lets non-programmers visually understand what “agentic” really means in their own context. ↩

Article
  { slug: "agents-need-their-own-spreadsheet"
  , originalUrl: (Just "https://synthesislabs.nl/insights/agents-need-their-own-spreadsheet/")
  , title: "Agents need their own spreadsheet"
  , date: constructDate 2026 9 30
  , keywords: ["agents","knowledge work","Kyyn","typed records"]
  , cover: { alt: "Tom Wells with the Synthesis logo", caption: Nothing, src: "/images/articles/agents-need-their-own-spreadsheet/cover-8115fdb89c53.webp" }
  , summary: "Agents don’t need more context. They need a workplace."
  , author: { bio: "Towards an AI-first, hyper-scaled, functional reactive universe!", email: "tom@tomwells.org", name: "Tom Wells" }
  , body:
      [ FlowLink { link: "https://synthesislabs.nl/insights/agents-need-their-own-spreadsheet/", title: "Originally published at Synthesis Labs · 30 September 2026" }
      , FlowImage { alt: "Tom Wells with the Synthesis logo", caption: Nothing, src: "/images/articles/agents-need-their-own-spreadsheet/cover-8115fdb89c53.webp" }
      , FlowRichParagraph [ InlineStrong "Agents don’t need more context. They need a workplace." ]
      , FlowRichParagraph [ InlineText "For the last few months I’ve been throwing agents at pretty much every bit of knowledge work I can find - sales reporting for our exco, figuring out who attended which training sessions, hunting down unused software licences (in addition to my day-job, where I’ve committed over 4,000 contributions - ", InlineLink { link: "https://github.com/drshade", title: "github.com/drshade" }, InlineText " for proof!). And honestly, the results are often amazing. An hour with an agent and you have something that would have taken a person a week." ]
      , FlowParagraph "But then next month comes around."
      , FlowParagraph "You open a new session, explain the problem again (a bit differently this time, because who remembers exactly what they said), paste in the new data, and get something… different. Different groupings, different caveats, a different total. Neither version is obviously wrong - but nobody can tell you why they differ, or which one to trust."
      , FlowParagraph "The work was great. It just lived on a whiteboard - and at the end of the session somebody wiped it."
      , FlowSection "Who wiped the whiteboard?"
      , FlowParagraph "Nobody, really. That’s just what a chat session is. Everything the agent knows sits in its context window, nothing it produced was checked against any kind of structure, and “run it again” actually means “ask again and hope”."
      , FlowParagraph "So you get clever and just keep using the same session. Month after month, same chat, same agent, all that lovely context still in there. Until one day, halfway through something important, you see the dreaded “conversation compacted” - and the agent wipes the whiteboard for you :)"
      , FlowParagraph "Chat is a great place to think. Keeping what you’ve thought is another matter."
      , FlowSection "A new “spreadsheet”"
      , FlowRichParagraph [ InlineText "Spreadsheets are not glamorous (sorry, finance friends), but they gave knowledge workers something we now completely take for granted. The rows and columns say what the data ", InlineEmphasis "is", InlineText ". The formulas say how it’s derived. Next month you drop in new numbers and the whole thing recalculates - the same way, every time. And when you move on, the sheet stays behind for someone else to open, follow and argue with." ]
      , FlowParagraph "Here’s the bit I love. When the spreadsheet arrived in 1979, accountants never learned to program - they learned to model. The tool asked for a new skill, and that skill ended up defining the next forty years of knowledge work. I think we’re at exactly the same point again. Teams won’t learn to code for their agents. They’ll learn to name their work precisely: the kinds of things, the fields, the links and the rules their agents operate on."
      , FlowParagraph "My strong belief is that agents need their own spreadsheet. Not a longer context window - a workplace. Somewhere with structure, memory and a review step, where the work survives the session."
      , FlowRichParagraph [ InlineText "And it matters more now than ever. Investors talk about ", InlineEmphasis "beta", InlineText " and ", InlineEmphasis "alpha", InlineText ". Beta is the return you get just for showing up - buy the whole market and everyone gets the same. Alpha is the extra you earn from knowing something nobody else does. Agents work exactly the same way. Frontier models already know the beta - the playbooks, the public docs, the best practice that every model ships with. What they don’t know is your alpha: why the deal actually closed, which supplier always slips, what the customer really meant. That’s the edge only your business has, and it decays fast and hides in inboxes and people’s heads. An agent with nowhere to keep it can only ever do the beta." ]
      , FlowSection "Carve it out first"
      , FlowRichParagraph [ InlineText "This doesn’t mean taking agents out of the loop. It means changing what the conversation is ", InlineEmphasis "for", InlineText ". The most valuable hour you can spend with an agent is carving out the structure: what are the things we care about, how do they relate, what are the rules, and where does the evidence come from?" ]
      , FlowParagraph "And this matters MUCH more than it sounds. While building that exco sales report I ran into a CRM rule that decided which division owned each deal. The rule was too blunt - one project spanning five business units landed entirely on a single division. No agent, however clever, could have reported that correctly, because the structure underneath was wrong. The fix was in the structure, not the prompt."
      , FlowParagraph "Get the structure right and the agent’s job gets much smaller. And much more reliable."
      , FlowSection "Enter Kyyn"
      , FlowRichParagraph [ InlineText "This is exactly the problem we’re building ", InlineLink { link: "https://kyyn.io", title: "Kyyn" }, InlineText " to solve. It’s in closed beta and we’re still testing it hard (so go easy on us!), but the idea is simple: give agents a proper place to do knowledge work." ]
      , FlowRichParagraph [ InlineText "A Kyyn knowledge base is a Git repository of typed records that carries its own schema - the rows and columns, if you like. The agent reads evidence from wherever it happens to live - mail, meetings, documents, Salesforce, finance systems, Git commits, a tidy structured store or a messy folder of dumps (it really doesn’t matter) - and ", InlineEmphasis "proposes", InlineText " changes. A human reviews the actual diff, and only accepted facts make it in, each one traceable back to where it came from." ]
      , FlowRichParagraph [ InlineText "Once the structure and the process are carved out properly, they stop needing you in the chair. That sales report now rebuilds itself every day from the CRM, the billing schedule and the budgets, with the same rules every time. Nobody has to remember what prompt they typed last month.", InlineFootnoteRef "1" ]
      , FlowSection "Slow to build, fast to run"
      , FlowRichParagraph [ InlineText "There’s a nice split of labour hiding in here too. Carving out the structure is slow, deliberate thinking - which is exactly what large language models are good at in conversation. But once it exists, most of what’s left are small, well-posed questions: ", InlineEmphasis "did this person actually attend the training, or were they only invited?" ]
      , FlowParagraph "Those don’t need a big model thinking out loud. They suit fast models that return a calibrated probability, with the threshold written into the process instead of buried somewhere in a prompt. Slow thinking to build the spreadsheet. Fast thinking to fill in the cells."
      , FlowParagraph "So here’s my question for you: how much of the work your agents did last month could you run again today and get the same answer?"
      , FlowParagraph "If the honest answer is “ummm, not much” - I’d love to chat. Kyyn is in closed beta, and real, repeatable knowledge work is exactly what we want to throw at it."
      , FlowSection "Footnotes"
      , FlowFootnote "1" [ InlineText "There’s some other fun stuff brewing in Kyyn too (still very much work in progress). We’re adding support for both ", InlineEmphasis "System One", InlineText " and ", InlineEmphasis "System Two", InlineText " style AI models - fast judgement models like ", InlineLink { link: "https://docs.typesafe.ai", title: "Jev" }, InlineText " that return a calibrated probability, alongside the big LLMs that reason slowly and deliberately. And we’re building agentic flows as type-safe, composable blocks, so you can see a whole flow before it ever runs. That lets non-programmers visually understand what “agentic” really means in their own context." ]
      ]
  }