--- title: "Build a Tiny Agent Studio" summary: "Lessons from making @howdy.notes: why one giant agent becomes context soup, and why useful AI work needs skills, feedback, randomness, and a half-code half-LLM production line." canonical: "https://howdynotes.com/ideas/build-a-tiny-agent-studio" type: "idea" status: "polished" created: "2026-08-04" updated: "2026-08-04" tags: ["agents", "ai-workflow", "systems", "creative-workflow", "howdy-notes", "automation"] ---

The first mistake is believing the agent is magic.

The second mistake is putting your entire life, brand, memory, task list, image style guide, posting workflow, and emotional support goblin inside one agent and asking it to “just know.”

That is not intelligence. That is a soup pot with Wi‑Fi. @howdy.notes got better when I stopped treating the agent like one giant brain and started treating it like a tiny studio. One brain. Many paws. The main agent is the editor. It holds taste, direction, approval, and judgment. The worker agents are paws. One scouts. One drafts. One checks crops. One remembers style feedback. One runs the boring code that refuses to post a chopped meme. This sounds less magical. Good. Magic is usually where bugs hide.

Bad model

One giant agent remembers everything.

Brand voice, image rules, posting logs, crop failures, old jokes, private preferences, every correction, every task, every vibe. All inside one haunted context window.

Better model

A tiny agent studio knows where truth lives.

The brain edits. The paws execute. Durable files hold rules. Code gates catch the stupid mistakes before they become public.

## 1. One giant agent becomes context landfill At the beginning, the tempting move was obvious: > Put everything in the agent. Brand voice. Workflow. Old mistakes. User preferences. Image prompts. Instagram rules. Meme references. Posting logs. Chinese caption notes. Crop failures. Random jokes that worked once at 1 AM. Then ask it to create something tasteful. This fails for a boring reason: attention is finite. No agent can hold infinite context perfectly. If you keep adding “important” things without structure, the important things become background noise. The result is not a wise assistant. It is a cat buried under 900 sticky notes, confidently sending the wrong PNG.

Memory is not a pile.

Memory is an interface.

Good agent work is not “more context.” It is: - the right files - loaded at the right time - for the right job - with hard rules where forgetting is expensive The agent should not remember everything. It should know where the truth lives. ## 2. Skills are not decoration. Skills are studio equipment. A useful agent needs more than prompts. Prompts are vibes. Skills are tools with handles. For @howdy.notes, the important shift was turning repeated taste corrections into durable procedures: - meme-generator rough first - then tricolor-cat redesign - preserve meme recognizability before style - separate English and Traditional Chinese posts - never treat a clean SVG proof as final art - safe-export before Instagram so the text does not get guillotined like a tiny content criminal Each rule sounds small. Together, they become a production system. Without skills, the agent keeps relearning the same lesson through public embarrassment. With skills, the studio gets better equipment. Not “be creative.” More like: > Here is the saw. Here is the ruler. Here is the sign that says DO NOT CUT OFF THE CAPTION AGAIN, YOU LITTLE MENACE. ## 3. Feedback is the actual training loop The most important dataset was not the internet. It was critique. “Looks like MS Paint.” “Good art style.” “White text on pale yellow? wtf???” “Quite fun haha.” “This is quite stupid.” Every one of these is gold, if it gets written down and changes the next run. Positive feedback matters too. Not in a motivational poster way. In a taste-calibration way. When something works, capture the mechanism: - Was it stupid-specific? - Was the meme readable in one second? - Did the silence do the stabbing? - Did the Taiwanese phrasing sound human instead of translated by a tax form? A good agent studio needs a Tone Vault, not just a prompt. Because taste is not a paragraph. Taste is a trail of accepted and rejected decisions. ## 4. Randomness is useful, but only inside a fence If the system becomes too consistent, it becomes stale. The same three meme formats come back wearing tiny fake mustaches. Drake again. Two Buttons again. This Is Fine again, somehow holding a different spreadsheet. So randomness matters. But pure randomness is just chaos with a confident caption. The useful version is fenced randomness: - choose from a weighted set of meme formats - avoid recent fatigue - force-connect the format to a real insight - kill it if the connection is weak - keep the editor brain in charge

Randomness is not the driver.

Randomness is the weird intern who brings in a cursed idea that might save the meeting.

Sometimes the intern is right. Sometimes the intern should not be allowed near typography. ## 5. The best system became half code, half LLM The LLM is good at messy translation: - turning a brain dump into a sharper thesis - finding the joke pressure point - creating variations - making the cat sound alive - explaining why a draft feels dead Code is good at being annoying in exactly the same way every time: - check dimensions - refuse unsafe exports - find newest relevant files - run preflight - verify expected package contents - prevent posting without approval This is the real architecture. Not AI replacing software. AI and software bullying each other into reliability.

LLM job

Find the joke.

Compress the messy idea, generate variations, explain why something feels dead, and keep the cat voice alive.

Code job

Say no.

Check the crop, validate the package, preserve the safe margins, and refuse to publish when approval is missing.

A fully manual workflow forgets. A fully automated workflow scales slop. A half-code, half-LLM workflow gives the cat a checklist and a knife. ## 6. The main agent should be the editor, not the factory The biggest improvement was separating roles. The main agent should not personally do every tiny operation while also holding taste, context, approval, scheduling, QA, and emotional damage. That is how you get overcooked assistant soup. The main agent should behave like an editor-in-chief: - decide what matters - protect the voice - assign work to paws - review outputs - reject slop - keep public actions approval-gated Worker agents should behave like studio crew: - read the durable files - produce the package - report clearly - disappear without becoming the main character This is not glamorous. It works. A tiny agent studio beats one giant haunted assistant because the work has shape. ## Picture notes for the illustrated version This article wants pictures, not corporate diagrams. Think ugly-cute tricolor cat field notes:
Panel Picture Words
1 A tiny chaotic studio inside a shoebox. Editor cat at the desk. Worker paws carrying tools. Don’t build one giant agent. Build a tiny agent studio.
2 A cat buried under 900 sticky notes, confidently holding the wrong PNG. More context can become soup.
3 A filing cabinet labeled style, posting rules, approved vibes, crop failures, current truth. The agent should know where truth lives.
4 Workshop bench with meme rough, cat wrap, safe export, hashtags, post approval tools. Prompts are vibes. Skills are handles.
5 A cursed feedback arcade machine eating critique tickets and outputting better drafts. Feedback is the training loop.
6 A goblin-cat intern shaking dice inside a fenced playpen while QA watches. Randomness helps. Unfenced randomness bites.
7 Split desk: LLM cloud-cat makes jokes; code-cat stamps REFUSED: UNSAFE CROP. LLM finds the joke. Code says no.
## The field note Do not build one giant agent. Build a tiny agent studio. Give the brain taste. Give the paws tools. Give the memory a filing cabinet. Give randomness a fence. Give code the right to say no. Then let feedback make the whole thing less stupid, one draft at a time. — howdy 🐱