DEV Community: Sanjay Shankar

AI Company Brain: How to Build One in 5 Steps

Sanjay Shankar — Wed, 12 Aug 2026 07:56:07 +0000

Quick answer: An AI company brain is a dedicated agent instance that lives in your team's chat, runs daily standups, tracks progress toward your goals, and generates reports automatically. I built one for my startup, named it Dojo, and open-sourced the entire kit. This guide covers the architecture, the 5 setup steps, and how to run it for your own team.

I first saw the idea in Antiwork's Gumclaw, the agent that now runs all of Gumroad's support and operations. It stuck with me: if an agent can run a company's day-to-day operations, it can at least run our standups. So I built our own version. This post is how it works, what it costs, and how you can have the whole thing free.

The Problem: Team Ops Was Scattered

Before Dojo, our founding team ran operations the way most small teams do: standups happened in one channel when people remembered, GTM tracking lived in spreadsheets nobody updated, and weekly reports took an hour to compile and were read by no one.

Pain Point	What It Cost Us
Standups were voluntary	No daily rhythm, no accountability
Metrics in spreadsheets	Numbers were always 2 weeks stale
Reports took manual effort	Nobody had time to write them
Targets felt abstract	No one could answer "are we on track?"

I wanted a system where the ops loop ran itself: the agent reminded, collected, compiled, reported, and tracked. Humans just answered and reviewed.

What a Company Brain Actually Is

A company brain is not a chatbot. It is an agent instance with a job: it knows your team, your targets, and your product; it lives in your chat; and it executes workflows on a schedule.

The architecture looks like this:

Always-on machine (we use a Mac mini)
  └── Hermes Agent (dedicated profile)
        ├── Discord gateway (team channel)
        ├── SOUL.md (identity, always loaded)
        ├── Skills (loaded on demand)
        │     ├── daily / weekly / monthly / quarterly workflows
        │     ├── meeting notes, support triage, sales leads
        │     ├── project tracking, onboarding
        │     └── roadmap, team, company knowledge, self-learning
        ├── Cron jobs (automated triggers)
        └── Data folder → JSON files → Dashboard HTML

The key design choice: everything is files and schedules, not magic. Daily standups compile into JSON, weekly reviews read those JSON files and generate a dashboard, and monthly reports read the weekly files. The whole system is inspectable and debuggable.

Step 1: Set Up the Agent on an Always-On Machine

You need a machine that stays on. A Mac mini, a spare laptop, or a small server all work. We used a Mac mini already sitting in the office.

Install the agent and create a dedicated profile for the brain. Giving it its own profile keeps its memory, skills, and config separate from anything else on the machine.

Pick a cost-efficient LLM as the provider. We run ours on DeepSeek, which keeps the monthly API bill low. The brain is not doing heavy reasoning — it is parsing standups and compiling reports — so you do not need a frontier model.

"The brain is not a genius. It is a reliable operator. Cheap models are enough when the workflows are well-defined." - Me, after the first month of running Dojo

Step 2: Connect It to Your Team's Chat

The brain lives where the team already talks. We use Discord.

The setup takes about 15 minutes: create a bot in the Discord Developer Portal, generate an invite URL with the right permissions, and add it to your server. The permissions matter more than people expect. The brain needs to read messages, send messages, read history, and mention people. Without mention permissions, your standup reminders are silent.

Create three channels so the brain has clear lanes:

Channel	Purpose
`#standup`	Daily standup input and reminders
`#reports`	Weekly, monthly, quarterly reports
`#alerts`	Nags, escalations, error alerts

Channel structure is how the brain learns where things happen. Consistency matters more than the specific names.

Step 3: Load the Skills (the Brain's Job Description)

Skills are the brain's playbooks. Each one defines a workflow: triggers, steps, parsing rules, and output format. Our kit ships 13 skills across three categories:

Category	Skills	What They Do
Meetings	daily, weekly, monthly, quarterly workflows + meeting notes	The full cadence: standup to strategic review
Operations	support triage, sales leads, project tracking, onboarding	The day-to-day work beyond meetings
Knowledge	roadmap plan, team, company/product, self-learning	What the brain knows, kept current

The self-learning skill is the sleeper feature. It watches conversations for signals like "X joined the team" or "project Y shipped" and updates the team and project files automatically, with confirmation. The brain maintains itself instead of rotting.

Step 4: Configure the Schedules

Cron jobs drive everything. The daily standup fires on weekdays, the weekly review on Friday afternoon, the monthly report on the 1st, and the quarterly review on the 1st of January, April, July, and October.

Time zones matter. All schedules run on your team's local time, so the standup reminder lands when people are actually at their desks. This is the kind of detail that separates a system people use from a system people ignore.

Step 5: Create the Data Folder and Baseline

The brain needs a place to store what it learns, and a baseline to compare against. Create the data folder structure, then add a baseline JSON with your team, your current numbers, and your targets.

This baseline is what makes the reports meaningful. "We added 120 leads this week" is a number. "We added 120 leads against a target of 140" is a signal. The whole reporting loop is built on that comparison.

What the Daily Loop Looks Like

Here is the rhythm once everything is running:

Morning - the brain posts a standup reminder with a simple format: Yesterday / Today / Blockers / Metrics
All day - team members reply in the channel; the brain parses each reply into structured data
Every 30 minutes - anyone who has not checked in gets a gentle nudge ("Even 'nothing for today' works!")
PTO handling - "off today" or "@brain off tomorrow" are understood natively, no nagging
Evening - the brain compiles the day's JSON and posts a summary

For our GTM tracking, the metrics are demos booked, leads added, and proposals sent. Yours will be different — that is why the kit keeps metrics as placeholders. The system works for any numbers you track.

Quick question: "What if nobody answers the standup?"
Answer: The brain nags politely, then escalates. Three days of silence from a project triggers an alert to leadership. The system does not force participation, but it makes silence visible.

The Dashboard: Where Data Becomes Decisions

The weekly workflow generates a self-contained HTML dashboard with charts: targets versus actuals, per-person contributions, trend lines, and a filter bar. It is generated from the JSON files, so it is always current.

We expose ours through a tunnel so the whole team can check it from anywhere. The dashboard is the answer to "are we on track?" that used to require digging through spreadsheets.

What It Costs

This is the part people ask about most. Our numbers:

Item	Cost
Machine	Mac mini we already owned (a used one is ~$300-500)
LLM API (DeepSeek)	$2-5/month for our usage
Discord	Free
Total ongoing	~$2-5/month

The brain runs standups, weekly reviews, monthly reports, and GTM tracking for less than the cost of one team coffee run. That is the whole pitch.

Challenges and How to Fix Them

Challenge: The brain misses replies or parses them wrong.
Example: A teammate writes "yesterday I did demos" without numbers, and the brain logs zero.
Fix: The parsing rules are explicit and documented. If a metric is ambiguous, the brain asks "how many?" rather than guessing. Corrections also feed the self-learning loop, so the same mistake rarely happens twice.

Challenge: People stop replying to standups after two weeks.
Example: Participation drops from 100% to 60% by week three.
Fix: This is a process problem, not a tool problem. The brain's nags help, but the real fix is leadership modeling: founders answer first, every day. A standup system only survives if the people at the top treat it as non-negotiable.

Challenge: The reports feel repetitive.
Example: The weekly report says the same thing three weeks in a row.
Fix: That is information, not a bug. If nothing changed, the team needs to ask why. The quarterly workflow adds strategy and hiring checks precisely so the cadence escalates in depth as the period grows.

Challenge: The brain goes quiet and nobody notices.
Example: The gateway drops and standups stop for a day.
Fix: The alerts channel catches this. Add a health check so the brain itself is monitored. An ops brain that silently dies is worse than none.

Frequently Asked Questions

Do I need a Mac mini to run a company brain?

No. Any always-on machine works: a spare laptop, a small server, or a Raspberry Pi class device. The brain is not compute-heavy. What matters is that the machine stays on and connected.

How long does setup take?

About 2-3 hours the first time: machine setup, Discord bot, skills, cron jobs, and a dry run. The kit's README walks through every step. A second install takes under an hour.

Which chat platforms are supported?

The agent works with Discord, Slack, Telegram, and others. We run ours on Discord. The kit's commands and workflows are written for Discord but adapt easily.

Is my team data safe?

The brain stores data as local JSON files on your own machine. Nothing is sent to a third-party SaaS. You control the machine, the data, and who has access. The LLM API calls carry only the specific prompt text.

Can I use a free or cheap LLM?

Yes. We run on DeepSeek for $2-5/month. The workflows are well-defined, so a cost-efficient model performs fine. You can upgrade to a stronger model later if needed.

What if my team uses different metrics?

The kit uses placeholders for metrics, targets, and team members. You replace them with whatever your team tracks: demos, sales calls, support tickets, engineering velocity, anything.

How is this different from a chatbot in our channel?

A chatbot answers questions. A company brain executes workflows on a schedule: it collects standups, compiles reports, generates dashboards, escalates blockers, and updates its own knowledge. The difference is the loop, not the chat.

Conclusion

A company brain turns team operations into a running system: standups that actually happen, targets that are actually tracked, and reports that compile themselves. We built ours for $2-5 a month, and the entire kit is open source on GitHub. Start with the daily standup workflow, run it for two weeks, and add the weekly and monthly layers once the rhythm sticks. The direction of the industry is clear — YC just open-sourced QM, a multiplayer harness for agent work. Company brains are becoming a category, and you can build yours this weekend.

Next step: clone the Company Brain Kit, follow the 10-step README, and run your first standup tomorrow. If you are building agent-driven teams, see how we think about the org chart model at visibility.so.

Found this useful? Share it with a founder who still compiles standups by hand. Follow @sanjayshankarr for more on running teams with AI agents.

Agentic SEO Services: What They Actually Include

Sanjay Shankar — Tue, 28 Jul 2026 05:39:25 +0000

A legitimate agentic SEO service should include four things you can verify: continuous automated audits, autonomous monitoring that opens and tracks fixable tasks, content agents with human approval gates, and AI visibility tracking across ChatGPT, Gemini, and Perplexity.
"Agentic SEO" turned into a category almost overnight. Every agency now has the word on its homepage. Some earned it. Most didn't.

What Real Agentic SEO Includes

1. Continuous Agent-Run Audits

Traditional audits are quarterly PDFs. Real agentic audits run weekly, diagnose issues, deduplicate against past findings, and create tasks automatically.

2. Autonomous Monitoring That Acts

A rank tracker that emails you is monitoring. An agent that notices a drop, checks competitors, drafts a fix, and assigns it — that's autonomous monitoring.

3. Content Agents With Human Approval

Real content agents structure for AI retrieval: entity definitions, schema markup, Q&A formatting, and heading-as-answer structure. Humans approve before anything ships.

4. AI Visibility Tracking

Real tracking runs actual prompts against ChatGPT, Gemini, Claude, Perplexity on a schedule and records whether your brand shows up.

5. Off-Page AI Trust

Over 80% of LLM citations rely on third-party web consensus. Entity consistency, Knowledge Graph, and citation gap monitoring matter as much as on-page work.

The 7-Question Vendor Checklist

Can I see a live audit run, not a sample report?
What triggers a task, and who approves it?
Which AI platforms do you test on, and how often?
What happens when the agent is wrong?
Does the pipeline learn from editor rejections?
Can I own the data if I cancel?
Is pricing tied to what's running or flat?

Full article with comparison table and detailed checklist: https://sanjayshankar.me/agentic-seo-services-what-they-include/

AI Product Manager vs Technical Program Manager: 5 Key Differences

Sanjay Shankar — Fri, 24 Jul 2026 13:12:34 +0000

I've been introduced as a Technical Project Manager, a Program Manager, and more recently a co-founder building AI products. Same person, same decade-plus of shipping software. The title kept changing because the work underneath it kept changing.

In 2026, that shift has a name: AI Product Manager. And a lot of Technical Program Managers are asking the same question — is this a new job, or is it just my job with a new label?
The honest answer is: it's both. Here are the 5 key differences.

What a TPM Does

A TPM's job is certainty. You take a roadmap that fifteen stakeholders agree on in theory and make sure it survives contact with reality — sprints, dependencies, scope creep.

What an AI PM Does

An AI PM owns questions a traditional PM never had to answer: Which tasks get delegated to an agent? What does "done" mean for probabilistic output? Where's the line between useful AI and AI theater?

5 Key Differences

Dimension	TPM	AI PM
Core question	Will this ship on time?	Should this exist?
Primary tool	Roadmaps, RAID logs	Evals, agent specs
Success measure	Delivery velocity	Adoption, trust
Failure mode	Blown timeline	Wrong output shipped
Where they sit	Between teams	Between model and market

The Builder-PM vs Integrator-PM Split

Builder-PMs prototype the thing themselves. Integrator-PMs evaluate vendor models and set policy. Neither is more senior — they're different bets on where you're most valuable.

Which Path Should You Choose?

Path 1: Stay in program management but add AI fluency. You'll be the person orgs trust to ship AI features safely.
Path 2: Push toward product. Your TPM background is an unfair advantage — most PMs can't scope technical risk like you can.

Full article: https://sanjayshankar.me/ai-product-manager-vs-technical-program-manager/

GEO vs AEO vs Agentic SEO: 3 Critical Differences Explained

Sanjay Shankar — Wed, 15 Jul 2026 05:31:59 +0000

GEO vs AEO vs Agentic SEO are three different answers to three different questions. Understanding GEO vs AEO vs Agentic SEO is essential if you are investing in AI search visibility.
Here is what each one actually means and how they work together.
I get some version of this question every couple of weeks now, usually in a LinkedIn DM: "Is agentic SEO just a new name for GEO?" or "Isnt AEO the same thing as what you are calling agentic SEO?"

No, no, and no. But I understand why people ask. All three terms showed up in the same 18-month window. All three are reactions to AI changing how people search. And marketing content being what it is, a lot of writers use them interchangeably to sound current.
They are not interchangeable. They answer three different questions about three different problems.

Quick answer:

GEO (Generative Engine Optimization) — will an AI chatbot mention my brand when it answers a question?
AEO (Answer Engine Optimization) — is my content structured so an answer engine can lift a clean answer out of it?
Agentic SEO — is the whole process of tracking, prioritizing, and fixing all of the above running on autopilot, or am I doing it by hand?

If you only remember one sentence: GEO and AEO describe what you are optimizing for. Agentic SEO describes how the work gets done.
Read the full article on sanjayshankar.me →

Understanding GEO vs AEO vs Agentic SEO: Three Different Questions.

1. GEO: Generative Engine Optimization

GEO is about visibility inside generated answers. When someone asks ChatGPT, Gemini, Perplexity, or Googles AI Overviews a question in your space, does your brand get named? Cited? Linked? Or does the answer skip you entirely and cite a competitor?
This is a genuinely new problem, not a rebrand of an old one. A page can rank number one on Google and still never get mentioned in an AI-generated answer. GEO is the practice of earning that mention: being clear, being consistent across the web, and being the source a model already knows.

2. AEO: Answer Engine Optimization

AEO is a content structuring discipline. It is about writing so that a featured snippet, a voice assistant, or an AI summary can pull a clean, self-contained answer out of your page without misquoting it.
In practice this means: a direct answer in the first sentence or two of a section, clear H2/H3 headers that match how people phrase the question, FAQ blocks, and schema markup.

3. Agentic SEO

Agentic SEO is not a content discipline. It is an operating model. It is the practice of using AI agents to run SEO as a continuous loop: monitor, find opportunities, turn them into tasks, execute, measure, repeat. Automatically, instead of waiting for next quarter's audit deck.

Comparison Table

	GEO	AEO	Agentic SEO
Core question	Does AI mention my brand?	Can AI extract a clean answer?	Who is doing the ongoing SEO work?
What you optimize	Brand presence in AI answers	Content structure and schema	The workflow: monitor, execute, measure
Where it shows up	ChatGPT, Perplexity, Gemini	Featured snippets, AI summaries	Everywhere above
Replacement for SEO?	No — layered on top	No — layered on top	No — it executes traditional + GEO + AEO

Which One Should You Focus On First?

Start with GEO if AI tools never mention your brand for questions you should own.

Start with AEO if you rank but keep losing the featured snippet to a competitor.

Start with Agentic SEO if you know what needs fixing but nothing ever gets done.

Full guide with examples and FAQ: https://sanjayshankar.me/geo-vs-aeo-vs-agentic-seo/

How We Built an AI Agent That Runs Technical SEO Audits Better Than Most Humans

Sanjay Shankar — Tue, 07 Jul 2026 06:18:46 +0000

Technical SEO audits should be automated. Not because humans are bad at them — because they're too slow.

A thorough technical audit of a mid-sized site (5,000–10,000 pages) takes a skilled SEO specialist roughly 2–3 days. By the time the report is delivered, the crawl data is already a week old. By the time the dev team acts on it, the findings are stale.

Tools like Screaming Frog and Sitebulb solved the crawling problem, but left the rest of the workflow untouched — downloading reports, analyzing issues, prioritizing fixes, assigning tasks, following up. Each step is manual. Each handoff loses context.

We built visibility.so's audit agents to solve one specific problem: close the gap between "here's what's broken" and "it's being fixed."

Full article on sanjayshankar.me →

The Problem: Audits Don't Scale

Most SEO teams follow a familiar pattern:

Run a desktop crawler — manual trigger, one-time export
Open the CSV in a spreadsheet
Manually prioritize issues by severity
Copy-paste findings into a task tracker
Assign tasks to developers via Slack or email
Follow up repeatedly
Re-crawl a month later to verify fixes

That's 7 steps for a single audit cycle. Half of them are administrative — not analytical.

Bottleneck	What It Costs You
Desktop-only tools	Can't schedule recurring crawls
CSV exports	Issues sit in files, not in your workflow
Manual prioritization	2–4 hours per audit cycle wasted
No task integration	~60% of audit findings never get actioned
No dedup across runs	Same issues reported week after week
No automatic follow-up	No way to verify fixes without re-crawling

The Architecture: Two Agents, One Pipeline

When you set a domain on any project in visibility.so, the system automatically provisions two audit agents. No configuration needed.

Agent 1: On-Page Auditor

Uses a real browser to inspect each page visually. It checks heading structure, meta tags, content quality signals, internal linking, image optimization, and schema markup.

Agent 2: Technical SEO Auditor

Focuses on infrastructure-level signals: crawlability, indexing coverage, Core Web Vitals (measured from actual browser render), structured data, performance, and security.

How an Audit Run Works (Step by Step)

Step 1: Scheduled Trigger

The scheduler wakes the assigned agent and passes context — project, domain, audit type, and existing issues list.

Step 2: Browser Access

The agent starts with a full browser session, allowing it to render JavaScript, measure Core Web Vitals, detect lazy-loading issues, and capture screenshots.

Step 3: Intelligent Crawling

Prioritized crawl: sitemap first → high-authority pages → canonical URLs → redirect chains. A 10,000-page site takes ~40-50 minutes.

Step 4: Issue Detection

Every finding is classified: critical, high, medium, or low — with suggested fixes.

Step 5: Dedup

The agent checks existing issues before logging. Same broken link? Already reported. Only new issues surface.

Step 6: Automatic Task Creation

Critical and high-severity issues create tasks in the project board with description, fix suggestion, and screenshot evidence.

From 3 Days to 47 Minutes

Time	Event
02:00 UTC	Scheduler triggers the agent
02:05–02:47	Agent crawls site with real browser
02:47	Agent logs 14 new opportunities
02:48	Agent creates 3 critical tasks
08:30	Team reviews findings
09:15	Dev team picks up first task

Agent time: 47 minutes. Human time: 45 minutes. Manual cycle: ~28 hours.

That's a 96% reduction in time-to-action.

Comparison

Capability	visibility.so	Desktop Crawler	Manual
Recurring schedule	✅	❌	❌
JavaScript rendering	✅	❌	✅
Task creation	✅ Auto	❌ CSV	❌ Manual
Dedup across weeks	✅	❌	❌
Time per audit	~47 min	~4 hrs	~2-3 days
Cost per audit	Included	£149/yr	~$800 billable

This article was originally published on sanjayshankar.me.