Good morning.

I recently ran a live briefing called The Asymmetric Operator for a small room of operators already building with agents.

Most of them came in trying to solve the same problem:

How to get fifteen agents to coordinate into a system instead of fifteen tools running in parallel.

Coordination is a real question. It is also the wrong place to spend the next twelve months.

Today's issue is the recap, cleaned up and rewritten for the newsletter. It covers where agent capability sits right now, why coordination will largely be solved at the platform level by end of year, what comes after that, and one thing you can install this week that compounds past any platform cycle.

— Sam

Let’s get into it.

Take a look at the agent market. 

General purpose agents: Manus, GenSpark, DeepAgent, Junior, Proxy, Computer, a dozen others. Sales and go-to-market agents: Rox, Piper, Ava, Sybil, Jordan, Mutiny, Omni. 

The same pattern repeats across customer support, research, coding, and every other category where operators deploy agents at scale. 

Each category has ten to twenty platforms sitting on top of the same underlying models, separated by a thin layer of differentiation.

A year ago most of these were demos. Now they are products. The capability question, can the model do the thing at all, is closed for almost every job you would put an agent on at your scale.

I know an e-commerce operator running about $650K a year. Him, three employees, a couple of contractors, roughly 80% of the operation sitting on top of one general-purpose agent platform. 

Every couple of weeks he asks whether he should switch to the new thing. The answer has been no every time. Double down on what works.

If the only thing you are taking from the last year of demos is that agents can do a lot, you are missing the thing that separates operators who compound from operators who stay flat. That thing has nothing to do with which tool they picked.

Every operator I talk to has landed on the same diagnosis, and the diagnosis is correct. The problem is coordination. 

Fifteen agents with fifteen different contexts, nothing shared, nothing compounding, the content agent having no idea what the sales agent saw last week. The instinct is to fix this by switching orchestration platforms. CrewAI to LangGraph to Claude Code to whatever launches next. It never works because the problem was never which tool.

Here is the honest twelve-month read: Coordination will largely be solved at the platform layer by end of year. 

The same agent platforms capturing your fix-it decisions, which you have been making all year every time an agent fails and you adjust it, have accumulated enough of that data to let agents self-correct.

When that ships, the operators whose only edge was "I got good at managing agent failures" lose the edge.

Which means building your whole position around coordination buys you nine months of advantage before the default tools catch up. 

You need to do the coordination work, and you should also assume the version that ships into the default tools inside twelve months is good enough for most of what you are hand-coordinating today.

The place to look is the layer underneath the agents themselves.

Cognitive architecture, at the file-and-folder level, is a small set of documents every agent in your business draws on. 

A CLAUDE.md file (or AGENTS.md, which is becoming the cross-model standard) holds the master context: business state, voice, decision boundaries, current priorities. 

A SOUL.md per agent holds that agent's identity, behavior, and hard limits. 

A skills folder holds modular instructions multiple agents can load on demand. 

A context folder holds business-specific knowledge like ICP profiles, offer details, and process documentation. 

A memory layer logs what each agent has learned across past runs. 

An agent team map describes who exists and how they connect.

Five to seven files, structured so the business has a single source of truth instead of fifteen agents each guessing at context from a fresh prompt.

Most operators building agents in 2026 have none of this. They have one agent's system prompt, maybe two, and a Notion page somewhere with "source of truth" written at the top that nothing reads from.

Set up properly, cognitive architecture is the reason some operators' agents compound while others stay flat with more tools. It is also only half the answer. The other half is what makes it cognitive.

Architecture is the files and folders. Cognition is the thinking an agent operates with inside that architecture.

The move almost nobody is making is capturing thought patterns and feeding them into the system. Thought patterns are your reasoning about the work your agents do, captured in writing. The path to the decision matters more than the decision itself.

One illustration: You run email marketing. At the end of the week you look at the last seven days: five campaigns sent, three subject lines pulled the open rate, click-through rate dropped across all five. You decide the problem is the calls to action. They read weak, with no urgency and no value in the click. You rewrite the CTA approach for next week's sends and log what you did and why, in about ten minutes. Voice-to-text works.

Do that every day, per agent cohort, for three weeks. You will have a body of captured reasoning that becomes the Thought Patterns section of that cohort's CLAUDE.md.

A few things people ask when I teach this. The first is whether the patterns need to be ideal. They do not. The volume and the feedback loop do the work. 

If you are capturing decisions over time and adjusting when results come back, the averages move up. 

You learn to make better decisions by logging the ones you make. 

The only way this fails is if you keep making bad calls on the business itself, and no cognition layer saves a business from that.

The second question is where to start. Begin with one agent cohort, usually marketing for most operators because the tasks repeat and the outcomes are measurable. Building the cognition layer for the entire business on week one is the same move that kills the habit by week three.

The third is what to do when the domain is not your strong suit. If you know you are weak at support or paid ads or whatever it is, the person on your team who is good at it captures the thought patterns for that cohort. Make it part of their work. Your lowest-judgment area is not where you want your agents learning from your decisions.

One of my clients, a B2B SaaS positioning consultancy, ran this build over six months. Four years old, founder plus three contractors, eight-week positioning sprints at $45K flat, $1.4M in revenue. Good business, flat trajectory. The founder could feel the ceiling without being able to name it.

Month zero: twelve GPTs doing pieces of the work, Perplexity for research, Claude for synthesis, Notion holding SOPs. Nothing shared context. Every engagement restarted the same orientation from scratch.

Month one through three: we built the cognitive architecture from the ground up. CLAUDE.md, SOUL.md per agent, eighteen skill files, context documents, and a memory layer, all in place. The team started capturing thought patterns weekly per cohort. By month three the reviewer agent was catching drift the founder used to catch on final review, and founder execution hours dropped by forty percent.

Month four through six: more agents, more patterns, more output. By month six the positioning hypothesis generator was scoring hypotheses against real client outcomes more accurately than the founder predicted them. The senior copywriter moved to founder-level taste work on the top 20% of engagements.

Six-month delta: engagement capacity from 25 a year to 52 a year run-rate, revenue from $1.4M to $2.1M at the same pricing, founder execution hours from 32 a week to 9, time to first strategic draft from three weeks to four days. NPS held flat, which is the number that matters most for the quality side of the story.

They did not build seven agents in a week. They built one, captured patterns, watched the outcomes, then added the next on top of what the first had taught them. The real asset is the six months of decisions and outcomes sitting on top of those agents. No competitor can buy that. They would have to live through the same six months of work to produce it.

Most operators using AI today are still Reactive. Humans do the work, agents answer one-off questions, every session starts cold.

A smaller group is Automated. Agents execute predefined work on schedules or triggers: weekly competitive scans, inbound lead qualifiers, content repurposers. Real leverage, for the slice of the business that is wired.

A handful are Attentive. Agents watch the business, interpret what they see, and surface what matters before the operator thinks to ask. Attentive requires a cognitive architecture underneath.

Almost nobody is Autonomous. Agents act inside authorized boundaries, learn from outcomes, and close loops the operator used to close. Autonomous compresses into real availability over the next twelve months for operators with the foundation in place and stays aspirational for everyone else.

Each rung requires the one below it, and the stack only holds if you build bottom-up. Attentive agents need a cognitive architecture to interpret from. A cognitive architecture needs a redesign of how the work is understood before you write a line of CLAUDE.md.

Three things to name about the next twelve months. Agent-to-agent protocols are hardening (MCP went from novelty to standard inside a year, and AGENTS.md is on the same trajectory). The dataset gap becomes unbridgeable: every month you are capturing thought patterns and outcomes, operators who start in six months will be six months further behind with no way to buy their way forward. And agents begin making strategic decisions instead of only tactical ones, which means operators without the architecture underneath can no longer see what their agents should do next.

This is the cheapest and fastest it will ever be to build this. A year from now the same position costs more to reach, and you reach it while competitors who started earlier already have the dataset compounding.

One thing to install this week, and the scope is small on purpose. One section of a single file, plus a ten-minute daily habit to populate it.

Step 1. Pick one agent cohort. Either one you already have running, or the first one you plan to build. Marketing is the easiest starting point for most operators because the tasks repeat and the outcomes are measurable.

Step 2. Open that cohort's CLAUDE.md, or create it if one does not exist yet. Add a new H2 section near the bottom titled Thought Patterns. Leave it blank for now.

Step 3. Each day this week, spend ten minutes logging the decisions you made about that cohort's work: what the data said, how you weighed the options, what you decided, and why. Voice-to-text is fine, and rough prose is fine. The goal is to capture the reasoning clearly, because the reader is a large language model that cares about context and not prose polish.

Step 4. Friday afternoon, look at the week's notes. Pull out the five to seven thought patterns that keep showing up across different days. The recurring ones are the signal.

Step 5. Paste those five to seven patterns into the Thought Patterns section you created in Step 2. Keep them as short, plainly-stated paragraphs. "When open rates drop below X, check send time first, then subject line, then audience segment" is the flavor.

Step 6. Repeat the next week. Extend the patterns, refine the wording, kill the ones that did not hold up. Over a month you will have a living thought-patterns file the cohort's agents can draw on, and your agents' outputs will begin to read like a sharper version of your own reasoning rather than a generic model's best guess.

That is the install. After the first pass it takes less than an hour a week, and it runs one cohort at a time. Do not try to do it for the whole business simultaneously; that is the exact move that kills the habit before it compounds.

The point of all of this is to build the part of your business nobody else can copy, and to put it in the files your agents read from.

Thought patterns are one of nine cognition types I walk operators through inside the advisory work.

A few of the other (memory loops, heuristic boundaries, taste calibration, and a few more) are going into Cortex across the next several issues. Signals covers the strategic case. Circuits delivers the build.

If today's issue landed, Cortex is where the work gets done.

And Cortex is open for new subscribers:

Cortex is a monthly premium version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.

You only get access to the Cortex (Signals and Circuits) issues for the months you're subscribed to, not the archives.

You can cancel at any time. But there are no refunds, for any reason. All sales are final.

Get in now in time for the next batch of issues.

Until next time,
Sam Woods
The Editor

.