Zac BrooksEight levels from a chat window to a business that shapes itself, and the blockers on the way.
The escape
The biggest gap in the market isn’t a tool. People who don’t think like systems builders never escape the chat window, so their AI never starts to compound.
- Chatting
Prompting back and forth gives back what you put in, one task at a time, and what comes out gets copied somewhere else. Nothing carries to the next task, so the value stays flat however good the model gets.
- Building systems
The business’s context in front of the agents, the work turned into skills and loops that act where the work lives, and every run building on the last. Slower at first, then it compounds.
You are here
Eight levels from a chat window to a business that shapes itself. Each starts with a challenge you’ll recognize, and the level is what it looks like solved. Each level opens below into its parts.
Chatting
AI as a chat window: useful in the moment, gone when the chat closes.
You type the same context into a chat window every time, and the good answers disappear when the chat closes.
What works is saved and reused.
This part is hard. Most don’t make it through because the context never leaves people’s heads, so each task starts from scratch.
The chat ends as a brainstorm or a doc, and someone has to carry it to the place it is actually used.
Agents act where the work lives: they publish, update and shape it directly.
A different AI tool for every task, none of them connected, with copying between them.
One place the work happens, connected to everything else.
The waterline: the escape. Most people stop above it.
Context
The company’s knowledge, live and permissioned, in front of every agent.
Every task starts with explaining the business again, and the answer is only as good as what you remembered to say.
The right facts and tools are already in front of the agent.
What the company knows lives in people’s heads, inboxes and old threads, where no agent can reach it.
Every meeting, email and document becomes memory agents start from.
The AI only knows what someone pasted in, and by tomorrow it is out of date.
Meetings, email, chat and tool events stream in as they happen, so the brain stays current.
The data sits in a dozen tools, some with connectors, some only as exports or dumps.
One set of connections feeds the brain.
Sharing the brain means everyone sees everything; locking it down means nobody gets the value.
What each person and each agent can see is set in one place, so the brain is shared safely.
This part is hard. Most don’t make it through because it feels all or nothing, and locking it down keeps the value from arriving.
Old notes and outdated decisions sit beside new ones, and the AI can’t tell which is true.
Stale information is pruned, and the newest wins.
This part is hard. Most don’t make it through because the answers slowly get less reliable, and trust in the whole system fades.
What we built for itConnections, brains and mining
Systems
Workflows turned into skills, agents and loops that anyone can start.
The workflow lives in one person’s chat history, and nobody else can run it.
It becomes skills, agents and loops anyone can start.
This part is hard. Most don’t make it through because knowing how to build one the AI-native way is still rare.
Working with agents means sitting in a chat, waiting on each step and babysitting it.
You start work, watch it run, and step in beside it.
Everything is built around one vendor’s model, so a better or cheaper one means starting over.
The model is a swappable part.
This part is hard. Most don’t make it through because a change of model can mean a rebuild, so it feels safer to wait.
What we built for itConnections, brains and miningAgent interaction
Trusted
Agents that hold a task for hours and come back with work you can use.
Agents lose the thread after a few minutes, so someone has to restart and steer them.
They hold a task for hours.
The same request gives a different result each run, so nothing can be relied on.
The same method gives the same result.
This part is hard. Most don’t make it through because checking every run takes back much of the time it saved.
All the effort goes into the agent that builds, and nothing checks its work before it ships.
Every maker has a checker.
This part is hard. Most don’t make it through because without a checker, a confident mistake can slip through and dent trust.
The output sounds like AI, not like you, and customers can tell.
It sounds like you.
This part is hard. Most don’t make it through because work that sounds and looks like AI is hard to put in front of customers.
The work arrives as Markdown or a plain doc that has to be rebuilt before anyone sees it.
It arrives designed and ready to send.
This part is hard. Most don’t make it through because rebuilding each output by hand stops being worth it.
What we built for itAgent interactionSkill trainingContent and brand engines
Running at volume
More work through the same system, with the cost of each run falling as volume grows.
Usage grows and the bill grows faster, until someone asks what it is really costing.
The cost of each run falls as volume grows.
This part is hard. Most don’t make it through because the cost of model spend grows faster than the value, and the pursuit gets reconsidered.
Every new request turns into a new build from scratch.
It plugs into the system you have.
A thousand agents produce work, and one person has to read and approve all of it.
A review layer shows what got done and takes approval in one pass.
This part is hard. Most don’t make it through because review becomes the bottleneck, and the work waits on one person.
The work agents now do at volume is still tracked in spreadsheets and chats.
A system built for this part of the business runs it.
Every video, deck and campaign is still made by hand, one at a time.
Engines make them on brand, at volume.
Agents can touch real data, and one wrong command can delete the only copy.
Data is protected, snapshotted and permissioned.
This part is hard. Most don’t make it through because one deletion or leak can be enough to end the experiment.
What we built for itSkill trainingContent and brand enginesThe review and system layerMarket maps and campaign infrastructureDeep research
Compounding
Every rep makes the system better and cheaper, and able to take on more.
The system only does what it was first built for, and each new problem starts again.
Each rep unlocks problems it couldn’t take on before.
Quality stops improving, because nothing feeds the results back into the method.
Every rep raises quality and hardens the method.
Each skill works on its own, and a person still carries the work between them.
Solid skills become steps of bigger ones, up to whole outcomes.
What you need to know is scattered across thousands of sources nobody has time to read.
Missions run across all of it, and the data becomes leverage.
Every internal tool you need becomes a software project with a queue behind it.
Specs go in, software comes out, and you watch the proof.
Having built something, you defend it, even when something better has shipped.
You burn it down and rebuild when something better arrives.
This part is hard. Most don’t make it through because what they built three months ago is already behind, and rebuilding feels like starting over.
What we built for itMarket maps and campaign infrastructureDeep researchThe software factorySkill chaining
Applied
The loop aimed at one business outcome, with a number that moves.
The systems are busy, but nobody can say what they earned or saved.
Every system ties to an outcome.
This part is hard. Most don’t make it through because the motion is easy to see and the results are hard to name, so support fades.
Everything is a priority, so nothing gets enough to move.
One gap on the wheel, one bet.
There is no before and after, so nobody knows if a change worked.
An OKR tied to revenue or cost, tracked.
One person gets ten times faster, and the rest of the company doesn’t.
The whole company runs on it.
This part is hard. Most don’t make it through because the gains stay with one person and are hard to spread.
What we built for itSkill chaining
Self-shaping
The system reads its own results and shapes the business, inside limits people set.
Every change waits on you to notice it, decide it and kick it off.
The system proposes and runs the next change.
Give agents room to act and they wander away from the goal.
Goals, limits and an off switch keep them on course.
This part is hard. Most don’t make it through because drift from the goal is hard to spot until a result comes back wrong.
With the system making changes, it is unclear who is steering.
People set the direction, and the system does the shaping.
What we built for itSkill chaining
The blockers you will run into
We ran into each of these building our own systems, and they are where we see people most often stall. You may meet some of them along the way, and knowing them early makes them easier to get through.
You may find yourself re-explaining the business to a chat window, with each task starting from scratch.
Stuck promptingThe access question tends to come early: sharing the brain widely feels risky, and locking it down limits the value.
All-or-nothing accessOver time, older context can creep into newer answers and quietly make them less reliable.
Context rotYou’ll know it should be a system, and your team may not yet know how to build one the AI-native way.
Building the systemYou may build around one model, then see a better or cheaper one ship a few months later.
Model lock-inResults may vary from run to run, and checking each one can eat into the time it saves.
RepeatabilityIt’s natural to put most of the effort into the agent that builds, and less into one that checks its work.
CounterbalanceThe output may sound and look like AI, and customers can often tell.
Sound and lookThe work may arrive as Markdown in Obsidian or a plain doc, and need rebuilding before a client sees it.
Markdown and plain docsAs usage grows, the cost of model spend may make you reconsider the pursuit.
Model spendAgents can produce more than one person can review, and the review can become the bottleneck.
The feedback waveOne wrong command might delete the only copy of something that mattered.
Lost dataYou may struggle to keep up with the cutting edge when what you built is already three months old.
Keeping upYou’ll begin to see motion and productivity, and still be thirsty for results.
Motion without resultsOne person may get much faster while the rest of the team works the way it always has.
One-person gainsAs agents get more room to act, they can drift from the goal in ways that are hard to spot.
Goal drift
What we had to build
Nine core competencies the market hasn’t solved. Each takes a session to explain, and each is an exponential gain, not a fix for one blocker.
01Connections, brains and mining
Lifts levels 1 to 2Everything the company knows, feeding in as it happens and mined into a memory every agent starts from.
02Agent interactionStarport
Lifts levels 2 to 3The layer where people and agents build together. Model providers solve about 90 percent of it, enough to sell more tokens, and almost every company lands on the rest unsolved.
03Skill trainingOrbits
Lifts levels 3 to 4Skills made repeatable and trained like an education, on frontier and open-weight models. It answers two risks at once: quality in execution, and the exponential cost of using AI.
04Content and brand engines
Lifts levels 3 to 4Pipelines that make on-brand work at volume. We have built them for video, for slides, and even for software.
05The review and system layerTerra
Lifts level 4When agents work, success arrives fast and brings two needs at once: a review layer to see and approve what got done, and a system built to run the workflow.
06Market maps and campaign infrastructureHarvista
Lifts levels 4 to 5Every company and person in a market, mapped, ranked and worked from one place.
07Deep research
Lifts levels 4 to 5Missions sent far and wide across the fragmented web, running a complex skill on every item they find. It breaks data moats and turns data into leverage.
08The software factory
Lifts level 5A spec goes in, and agents build, test and ship it. Ours keeps the lights on: you watch the proof play back instead of reading the code.
09Skill chaining
Lifts levels 5 to 7Once a skill is solid, with its own steps, it becomes a single step inside a bigger skill, and that skill a step inside a bigger one again. This is the work at the top of the market.
AI-native, wedge by wedge
At Level 6 the loop is aimed at the wheel. This is what each of its six dimensions looks like at the top of the path.
Research never stops. The market, the problems and the competitive picture refresh on their own, with every fact graded.
GrowthAgents find who to call and why, a person sends every message, and the pipeline keeps itself current.
ExecutionThe playbook is a trained method. Volume and variety arrive without hiring, and a second agent checks every step.
InformationNobody writes a status update. Everything is captured, searchable, and the first thing every run reads.
IterationEvery change carries a goal and a before and after, and each method gets cheaper every time it runs.
RiskSafeguards are built into the system, not written in a rule. The kill switch works, and agents never send on their own.