> ## Documentation Index
> Fetch the complete documentation index at: https://lab.pollack.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Agento Studio
> A systematic approach to growing AI agents — judges tell you if the agent worked, journals tell you why
Grow your agents, don't just prompt them.
Run. Judge. Read the journal. Fix the hotspot. Run again. That's the loop that turns unpredictable agents into reliable ones.
Agento Studio is a systematic approach to growing AI agents. A **judge** tells you if the agent got it right. A **journal** tells you why it behaved the way it did. You need both — without the judge you're guessing, without the journal you're tuning blind.
github.com/markpollack/agento-studio
## Agents Are Workflows
An agent isn't a magic black box — it's a workflow. Each step is either **deterministic** (build, lint, test, measure coverage) or an **AI step** (reason about an error, generate code, decide what to fix next). What people casually call "an agent" is usually just the AI step — one node in a larger pipeline.
```
fetch PR → rebase → detect conflicts → run tests → fix & retest → cleanup → build gate
[determ.] [determ.] [deterministic] [determ.] [AI step] [determ.] [judge]
│
┌────────────────────────┐ │ pass
│ version-pattern judge │◄───┘
│ [deterministic] │
└──────────┬─────────────┘
│
┌──────────┴─────────────┐
│ parallel AI steps │
│ assess-code-quality │
│ assess-backport │
└──────────┬─────────────┘
│
┌──────────┴─────────────┐
│ quality judge → report│
└────────────────────────┘
```
This is the actual [AgentWorks PR Review](https://github.com/markpollack/agentworks-pr-review) pipeline — seven deterministic steps, a judge gate, then parallel AI assessment only if the build passes. Most of the workflow never touches an LLM.
This is exactly the pattern Stripe arrived at independently with their [Minions system](https://blog.pollack.ai/stripe-convergence/) — 1,300+ PRs per week at \$1T scale. They call the pattern "blueprints": deterministic nodes interleaved with agent nodes. As Stripe's Alistair Gray put it: *"Blueprints combine the determinism of workflows with agents' flexibility in dealing with the unknown."*
The insight behind Agento Studio: **the workflow structure matters more than the model powering the AI steps.** Better prompts, targeted knowledge, and deterministic checkpoints consistently outperform model upgrades. That's the thesis — `knowledge + structured execution > model` — and the build-measure-improve loop below is how you prove it for your agent.
## How It Works
1. **Run** the agent on a real task
2. **Judge** the output — did it actually work? (correctness)
3. **Read the journal** — why did it behave that way? (behavior)
4. **Fix the hotspot** — better skill, better prompt, better tool
5. **Run again** — measure if it improved
Each lever you can turn — skills, knowledge bases, pre-analysis, steering hooks — gets validated through this loop, not assumed to help.
## What You Can Build
The methodology supports four project variants — each with its own feedback loop:
| Variant | Use when... | Feedback loop |
| -------------- | -------------------------------------------------------- | -------------------------------------------------------------------- |
| **Eval-Agent** | Building an autonomous agent with judge-based evaluation | Loss optimization — judges score each run, journal explains behavior |
| **Project** | Bootstrapping a library, service, or application | QA review loop with vision → design → roadmap |
| **Research** | Investigating a question or testing hypotheses | Vision ↔ research iteration, multi-roadmap pattern |
| **Steward** | Ongoing maintenance of a project or domain | Health monitoring — continuous, not convergent |
### Agents Built So Far
| Agent | What it does |
| --------------------------------------------------------- | ---------------------------------------------------------------- |
| [PR Merge](/projects/loopy) | Automated pull request review, merge, and pipeline orchestration |
| [Issue Classification](/experiments/issue-classification) | Categorize and triage issues (tested against SWE-bench) |
| [Code Coverage](/experiments/code-coverage-v3) | Generate tests to hit coverage targets across multiple projects |
| Liquibase → Flyway | Migration agent for database schema tooling conversion |
Each one goes through the same cycle: run on real tasks, judge the output, read the journal, fix the hotspots, run again.
## Blog
* [I Read My Agent's Diary](https://blog.pollack.ai/i-read-my-agents-diary/) — Markov chain analysis of agent tool-call traces across eval-agent experiments
* [Look Ma, No RAG!](https://blog.pollack.ai/look-ma-no-rag/) — How the knowledge layer works: routing tables, two KB types, federation, and health checks
## Try It
The repo includes slash commands in `.claude/commands/` — they're available automatically when you launch Claude Code from within the repo:
```bash theme={null}
git clone https://github.com/markpollack/agento-studio.git ~/agento-studio
cd ~/agento-studio
claude
# then: /forge-research-kb ~/my-research-kb "your topic here"
```
To use the commands from another project, add the repo as a context directory:
```bash theme={null}
claude --add-dir ~/agento-studio
```
The [getting-started guide](https://github.com/markpollack/agento-studio/blob/main/guides/getting-started-research-kb.md) walks through building your first research-partner KB end-to-end — five seed papers, 20 minutes, a working research agent at the end.
## License
BSL 1.1 — converts to Apache 2.0 on April 1, 2029.
