# What a One-Person AI Org Looks Like at 90 Days > A one-person AI org at 90 days: 94 agents, 27 workflows, 4M+ words of compounding knowledge. What the platform contains, what it produced, and how to start. - Canonical: https://hcd.ai/ai-in-practice/one-person-ai-org-90-days/ - Published: 2026-05-25 - Updated: 2026-05-25 - Author: Sasha Merzliakov (https://hcd.ai/about/) - Pillar: AI in Practice ---

A one-person AI org is a structured team of AI agents, skills, workflows, hooks and memory that a single person can operate. The purpose of this article is to share how powerful Claude Code is beyond what's previously been primarily an engineering conversation, and what 90 days of effort spent after hours can translate into. I will be sharing exactly what that looks like, what value that's created for me personally and professionally, and where I see the next 90 days taking me. For those that have not taken the plunge into this space, my aim is to inspire and show what can be done beyond a generic ChatGPT interaction.

How to Get Started

What I'm about to share might come across as seemingly out of reach, but looking back, I believe anyone can take this journey too. Like any journey it starts with taking that first step. Unlike linear journeys, this one is defined by 'the exponential', meaning the steps you take might've been fixed stride lengths in the past but now with AI those steps become leaps and bounds, and even then there are some 'steps' I've reflected on and can only describe as superhuman. That first step, get Claude Code. Yes there are other options out there, for example ChatGPT codex, but this article is not about comparing frontier models. You can't go too wrong investing the next 90 days in Claude Code and much of your learnings will be transferable. If you're starting from scratch Anthropic Academy have a range of free courses to get you going.

What a one-person AI org contains

Here is what mine looks like at 90 days, laid out across the dimensions of capability stack, infrastructure, projects, output, volume, velocity. This shape is not prescriptive, yours will be different based on your objectives, and it's something that changes daily.
Dimension Count Description/examples
CAPABILITY STACK
Agents 94 9 chiefs/exco + 85 specialists
Skills 128 Reusable processes, archive shows the iteration trail
Hooks 4 Session lifecycle + safety guards
Workflows 27 Named orchestration chains across workstreams
Workflow steps 171 Individual stages within each workflow
Memory engrams 173 Persistent rules and learnings the AI carries across sessions
Episodes 281 One per session or ~4 per day
Global memory docs 9 Exec-Summary, architecture, glossary, soul, patterns, etc
Scheduled tasks 32 Market intel, Email digest summaries, home telemetry
Python scripts 297 Prediction models, backtesting, signal detection, reporting
INFRASTRUCTURE
Cloud services 6 Hosting, tunnel, three broker servers, bot host
Local deployments 3 MacBook Pro, Geekom A8 Max 24/7, Samsung SmartTV
Smart home 6 Sigenergy battery, Ecowitt weather, Tapo plugs, Deco mesh
Storage 195 ~165 GB model weights, under 2 GB code and prose
Git repos 11 Local repos across projects, mirrors on GitHub
Network/sync 2 Tailscale private network, Syncthing folder sync
Local LLM models 36 ~135 GB image+video, 30 GB Ollama text
External AI APIs 11 Anthropic, Gemini, OpenAI, Perplexity, Brave Search
MCP servers 7 Gmail, Calendar, Drive, Playwright, Discord, Figma, Apify
Active domains 4 Brand, product, and dev surfaces
PROJECTS
Project directories 72 Across 3 buckets: AI Org, Clients, Personal
AI Org 32 Agent teams workflows, Agents SDK, autoresearch, strategy
Clients 21 Non commercial client demos across product and design
Personal 19 Home automation, creative tools, health, finance
Dashboards 15 Operational UIs for trading, AI org orchestration, design
OUTPUT LAYER
Apps 5 Mobile + web product apps, mostly private
Articles total 41 32 published, 5 draft, 4 backlog
Curated resources 305 Across 10 categorised sections on hcd.ai/resources/
Design systems 4 Brand ref, org system, two client systems
LinkedIn posts 12 All tied to articles
Reports 177 Analytics pulls, AEO/GEO insights, audits across all projects
Research 733K 35 topics, 270 files, 733K words
Websites 4 Primary brand, org secondary, two client estates
VOLUME
Code LOC 490,612 HTML 134K, Python 93K, JS 66K, JSON 58K, TS 49K
Code-to-prose ratio 1.10:1 Was 1.12:1, drifting toward parity
Grand total LOC 937,842 Lines of code, includes markdown files
HTML files 67 Generated static pages on hcd.ai
Image assets 98 AI-generated heroes, social cards, diagrams
Markdown lines 413,375 Articles, research, briefs, and project docs
Prose files 818 Research 270, episodes 281, memory 173, agents 94
Markdown words 4,082,911 Word count across the same .md files
VELOCITY
Commit window 71 Total days, git not added straight away
Active commit days 64 Of 71 elapsed days
Touch rate 90% Was 93% on day 60, elapsed outpacing active
Total commits 305 ~4.8 per active day

Value created, the "so what?"

The inventory above is what a one-person AI org looks like at 90 days, but ultimately, so what? **Compounding knowledge.** The AI org gets better at my specific work the longer I run it. Nothing resets between sessions. Last week's rejected phrasings, the briefs that landed, the shortcuts I take when I am tired, the AI org has all of it on file. Each session starts further forward than the one before. **Always-on operating environment.** The platform runs when I do not. The market intel digest is already in my inbox by the time I have made coffee on Monday, with the overnight reports underneath it and the resources scout's proposed new entries underneath that. **The foundations to scale AI agent teams.** Adding agents takes a fraction of the time of what adding the first one did. Voice rules, brief templates, the memory layer, the verification pass, the lint, the deploy gates, all there on day one for new agents to inherit. The focus now moves on how to do more work, more quickly and with higher quality and autonomy. **The ability to automate and parallelise work.** Inside a single session I might have a research agent working one part of a brief while a peer-reviewer scrutinises another. Scheduled jobs pull the reports I used to forget to run. Briefs draft themselves while I am doing something else. Work that used to wait for me does not wait anymore. **The learnings and perspectives gained along the journey.** The hands on execution has taken my knowledge to another level. The journey has also radically changed how I think about creating and delivering value and the future ways of working, spoiler alert, it does not involve adding AI to what we currently do, but letting AI re-imagine how we can more effectively do it.

Examples of work (screenshots)

The following 7 projects represent a small selection of work completed over the 90 days. Each is working capability I regularly use.

AI Org memory. An engram index is the routing layer for every memory the platform carries between sessions. Each entry is a one-line pointer into an extracted learning, a feedback rule, a reference fact, or a project state. This is what compounding knowledge across interactions looks like at an organisational scale.

A Brain dashboard view of the AI org's global memory layer. The left panel shows overview stats including 47 patterns, 30 failures, 47 stale entries, 82 agents, and a confidence distribution histogram. The centre panel shows the consolidation pipeline as five phases: the Engram Index router (MEMORY.md), Working Memory (live, session-start hook), Episodic Memory (captured at session end), Consolidation (daily job at noon), and Semantic Memory (durable long-term layer). The right panel shows the /save-sync skill detail with its three steps.
The Brain dashboard: engram index as router, working / episodic / semantic memory layers, and the consolidation job that promotes what the AI carries forward.

Orchestration workflow. An agent dynamically designs each workflow, assigns the relevant agent, and renders the workflow in a human readable format. Each stage shows its owner and each verification gate shows the condition that has to be true before the work moves on. The diagram is what the human sees when reviewing and verifying, and what a new agent inherits when it joins.

A WORK dashboard view of the Platform performance review workflow run on 27 April 2026. Left panel lists three workflows: Platform performance review, AI Org evaluation, Brain / memory evolution. Centre panel shows the workflow as a vertical chain: CEO approval, COS /mc-platform-review skill, Phase 1 Architecture pass (solutions-architect), Phase 2 Domain reviews running three code-reviewer instances in parallel for server / frontend / data, Phase 3 Consolidation (solutions-architect), Phase 4 Verdict + decision row (chief-digital). Right panel shows the code-reviewer step detail with its input scope, output, and handoff.
Dynamic workflow creation by Chief of Staff orchestration agent, 'human-readable' visualisation to show stages, agents, parallel branches and human verification gates.

Next best action modeller. A customer-base dashboard organised around the DIKW pyramid: data, information, knowledge, wisdom, with cross-cutting graph and query layers. One scaffold carries churn propensity, retention scoring, incentive targeting and next-best-offer. Non-analysts can ask the dataset in natural language and get back tangible AI powered insights to execute.

An NBA AI dashboard WISDOM view showing applied business plays for two datasets side by side. Left panel for Cell2Cell Customer Churn (51,047 rows, 58 features, 28.8% churn) shows two numbered amber-accent play cards: 01 Proactive retention scoring (monthly top-decile retention contact based on high churn risk and high CLV) and 02 Root-cause feedback loop (pair retention scoring with a monthly cohort report to GTM so product and pricing can fix root causes). Right panel for IBM Telco Customer Churn (7,043 rows, 21 features, 26.5% churn) shows three plays: 01 Contract upsell (push month-to-month customers onto 1-2 year contracts with a $5/mo discount), 02 Service bundling (free first-year Online Security or Tech Support, drops churn 20-30%), 03 Payment migration (move Electronic-Check payers to auto-pay with a one-time bill credit). Each panel ends with a cadence footer. Left sidebar shows the DIKW navigation: DATA, INFORMATION, KNOWLEDGE, WISDOM (selected), GRAPH, QUERY.
Input a dataset, extract insights, execute next best actions.

Local media creation tool. A custom media platform for image and video generation, built on top of a local ComfyUI stack running open source models such as Flux. The interface lets you pick a render mode, set aspect ratio and stylization, and generate against a styles + recipes library, all running offline on my MacBook. Additional capability includes style creation, asset management and ad format generation.

A custom MEDIA platform with an IMAGES surface (Images, Video, Renders, Library, Ads in the left nav; Styles and Recipes under Assets). Centre column shows controls for Render Mode (Photoreal LoRA, Flux Dev + 4-LoRA realism stack), Aspect Ratio (2:3 / 1:1 selected / 3:2 with a scale slider), Aesthetics (Stylization 3.5, Variety 0%), Style Modifiers, Number of Generations (1 selected through 4), and Speed/Quality (Relax / Fast / HD selected). Main canvas shows a prompt input running 'an incredibly neat and whimsical creature with dr seuss vibes...' with the generated 1024x1024 turquoise-and-orange creature preview, plus a RECENT grid of recent generations including robotic and natural-form pieces. Dark theme throughout.
Full image-and-video generation pipeline, running locally against open-weights models.

Prompt tool. A multi-vendor evaluation workbench for authoring a prompt, defining criteria and test cases, then running it across Anthropic, Google and OpenAI models. Per-row scores surface strengths, weaknesses and reasoning, and an auto-iterate loop hands the lowest-scoring rows to a "prompt doctor" that proposes a revised prompt and re-evaluates.

Prompt Tool interface with a three-column layout for authoring, evaluating, and iterating prompts. Left column shows the PROMPT panel with a code-editor input containing role + instructions + {input} placeholder. Middle column shows EVALUATION CRITERIA as a free-form list (response directly addresses the user's question, no filler or preamble, factually correct, appropriate level of detail), TEST CASES as a JSON array of input/expected pairs with three cases ready, and model controls (Claude Opus 4.7, target 9.0, 4 rounds, Run eval / Auto-iterate / Cancel buttons). Right column shows AVERAGE SCORE out of 10, RECOMMENDATIONS with add/edit/delete actions referencing prompt line numbers, and PER-ROW RESULTS with strengths/weaknesses/reasoning per test case.
Multi-vendor prompt evaluation, with an auto-iterate loop to improve performance.

Home automation. The Home dashboard replaces a decommissioned Home Assistant VM with a Vite + TypeScript + SQLite stack purpose-built around the actual telemetry I care about. Battery state, solar generation, weather and switched plugs feed into one view. Claude Code powers wholesale energy buying and selling decisions through Amber.

Home dashboard Battery view, built in Vite and TypeScript on a SQLite store. Top row shows live tiles for battery state of charge, solar generation against installed kWp, house load, and grid import/export, each labelled with its Modbus register. A 30-minute Amber wholesale buy and sell price forecast runs across forty-nine intervals. A Sankey diagram visualises today's energy flow between Solar, Battery, Grid, and Load. Below that are Modbus actuals over the day, a forward projected-usage simulation, a 24-hour dispatch plan with charge and discharge slot bars, and a dispatch model rules panel with toggles for Maximise earnings, Sell from existing battery, Buy-low / sell-high arbitrage, Load-shift, Top up over lunch, and Reserve minimum battery. Left sidebar lists Battery, Amber, Weather, Irrigation, Network, Devices, Cameras, Calendar, Costs, Rules, Scenarios, Design System.
Custom Home dashboard, built for the telemetry that actually matters to me. Click to view full height image including Projected Usage and Dispatch plan

Trading agents. Autonomous agents execute a strategy on paper accounts, with the Trade Journal dashboard rendering setups, executions and outcomes against the rules they were given. Real-world autonomous execution in a domain where being wrong has a price. The journal is where verification happens after the fact.

A LIVE CHARTS view in the trading dashboard showing XAUUSD across three timeframes side by side: 1H, 5M, and 1M, each with green and red A+ zones marked. Top toolbar shows broker tabs (PEPP, FTMO, BLUE), an instrument row (GER40, US30, NAS100, XAUUSD selected, BTCUSD, EURUSD, USDJPY, GBPUSD, AUDUSD), and a risk-management strip showing balance, risk %, volume, stop loss, R-multiple, and take-profit. Right panel is a COACH chat surface with MINDSET tab and STOP button, showing real-time analysis of the current price action including HTF level rejection, 5M reversal candle, 2M bear sequence, volume read, and current proximity to the rejection high. Header tag reads 'WAIT, 2M HAS NO A+/A ZONES'.
Agents dynamically created based on defined characteristics, backtesting performance, and evolutionary selection, tracked live in manual trading environment.

Key learnings

The build itself has been the education. Five things stand out as the lessons I would carry forward. **Starting with one agent.** The temptation when you see what a platform like this can hold is to design the full org chart upfront and then populate it, or download a bundle from GitHub. I went the other way. The first agent that handled one workstream end-to-end taught me more than ten scaffolded definitions would have. Every working agent since has shaped the org chart more truthfully than any pre-packaged bundle could. **The org chart can run ahead of the work.** Most agents were scaffolded but unrouted, meaning they existed as definitions but had not been wired into any workflow. I built faster than I integrated, unwired agents are still valuable, they could just be even more valuable. The fix is unglamorous, wire every new AI agent into a workflow before scaffolding the next one. Workflows also help in delivering more consistent outputs. **The bottleneck flips from generation to verification.** AI generates faster than any human can check. Once that happens, the rate-limiter on the platform is no longer "can the AI write this?" but "can I trust what it wrote?". Aside from human review, agent peer-review, lint, deploy preflights and the verification pass are not optional polish. They are the architecture that keeps quality alive at the speed of generation. Skip them and the output becomes a quality crisis. **Memory value only compounds if you build it deliberately.** The AI org learns not because it's saving conversations or creating handoff files, but through human-like memory consolidation. The session-capture hook, the consolidation pipeline, the engram index, the per-agent memory files, none of these are defaults. Memory systems are gaining increasing attention. The real question however is whether the system is learning or just recalling. **Workflows provide a safeguard.** A deploy curl-check used source filenames instead of frontmatter permalinks, and a dry-run rsync flagged three live articles for deletion before the mismatch was spotted. A LinkedIn scraping job triggered an overnight account restriction and was retired the same day. Neither was an AI failure. Both were workflow gaps. Wire the rules in before you scale the volume.

The next 90 days

Looking back I could never have imagined where I would've ended up and I feel that's going to be the case for the next 90 days too! I'd like to think this was a discovery phase and next starting to define and develop a direction to take, but with the constant advance of AI it's clear there will always be the need for exploration and discovery. This is where I feel my attention will be: **Self-evolving agents.** Each agent to this point has been carefully crafted. The next step is to extend the capability of my Chief of Staff and their team to not only monitor agent performance, but proactively and automatically evaluate, model and modify agents. **Agent team performance.** Current work orchestration includes a human verification component. The next step is to explore increasing autonomy within the team, agents taking on increasing responsibility for other agents, and ways of working and team topologies to optimise performance. **Autonomous learning.** Building on the AI org's memory system, and more than a scheduled heartbeat, what does always-on autonomous learning look like, what rate of compounding can be achieved, and what value can be derived and measured from the process. **The exponential.** Using AI we often default to how this can help with our existing workflow or process. Agent teams can plan and execute their own workflow, which is a significant paradigm shift. I'm keen to explore what that next paradigm beyond this is. {% faqs faqs %}