Claude Engineering Pod
egorttt/claude-engineering-pod
Agentes de revisão de código e geração de testes prontos para produção
Ver no GitHubDois agentes Claude em estilo produção para automação de engenharia: revisão de código e geração de testes. Extraído do fluxo diário de um time fintech, com suporte a tool use, estado e avaliação. Inclui exemplos rodáveis e harness de eval para medir concordância com revisores sênior.
Instalação
npx convez-ai egorttt/claude-engineering-pod1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de revisão de pull requests com comentários estruturados ancorados em linhas
- Geração automática de testes para cobrir casos extremos e semântica de schemas
- Redução de tempo rotineiro em engenharia e melhoria de previsibilidade de sprints
Veredito

Implementação real e solidamente testada de agentes Claude com tool use e avaliação. Código legível em 30 minutos, pronto para ser adaptado; melhor para times que já rodam CI/CD e querem embutir IA no workflow diário de engenharia.
README
A tradução para o português deste README ainda está na fila.
claude-engineering-pod
Two production-grade agents I run inside my engineering team of 14. Extracted to a clean, self-contained repo so anyone can read the code, run the demos, and steal what's useful.
- 🔍 code_review — a code-review agent with tool access to a diff, file reads, the repo's style guide, and (optionally) git blame. Produces structured, line-anchored review comments. Eval harness measures agreement with senior reviewers on a labelled fixture set.
- 🧪 test_gen — a test-generation agent that enumerates test categories from a piece of code, then writes runnable tests for each. Built to cover the unglamorous edge cases (chunk boundaries, ON CONFLICT semantics, schema drift) that humans skip when bored.
Both are agents with state, tool use, and evaluation — not one-shot prompts. They're written to be small enough that you can read all the code in 30 minutes, swap models in one place, and rip out what doesn't fit your team.
Why this exists
For two years I've been running both of these inside a 14-person fintech engineering team. We measured ~25% reduction in routine engineering time and +15-20% improvement in sprint predictability across two quarters.
The internal versions are tied to bank integrations and customer code; I can't open-source those. So this is a stripped-down, language-agnostic version of the same two patterns, sized to be useful as a starting point for any team that wants to embed agents into their daily engineering loop.
What's inside
claude-engineering-pod/
├── src/
│ ├── code_review/ review agent
│ │ ├── agent.py main loop
│ │ ├── prompts.py versioned prompt templates
│ │ └── tools.py diff parsing, file reads, style guide
│ ├── test_gen/ test-generation agent
│ │ ├── agent.py
│ │ ├── prompts.py
│ │ └── tools.py code analysis, test category enumeration
│ └── shared/
│ └── llm.py Claude SDK wrapper + tool-use loop (one screen of code)
├── examples/
│ ├── review_a_pr.py runnable example, ~50 LOC
│ └── generate_tests.py runnable example, ~50 LOC
├── evals/
│ ├── code_review_evals.py
│ └── test_gen_evals.py
├── fixtures/ small labelled dataset for evals
└── docs/
├── architecture.md
└── learnings.md what I'd do differently after 2 years
Quick start
git clone https://github.com/egorttt/claude-engineering-pod
cd claude-engineering-pod
python -m venv .venv && source .venv/bin/activate
pip install -e .
cp .env.example .env
# put your ANTHROPIC_API_KEY in .env
python examples/review_a_pr.py fixtures/example_diff.patch
python examples/generate_tests.py fixtures/example_module.py
That's it. Both examples are <60 lines each, fully runnable.
Design choices I'd defend
A few things that took us a while to figure out, that I'd argue for in any team adopting this pattern:
-
Prompts as versioned code, not config. Every prompt lives in
prompts.pynext to the agent that uses it. They get diffed in PRs, reviewed like code, and tagged when behavior changes. Treating prompts as separate config files in YAML files is how you lose track of which version of the prompt produced which output in production. -
Tools are the contract, not the implementation. Each tool has a strict JSON schema and a Python implementation. The agent doesn't see how
read_fileworks — only that it returns text. This means you can swap the implementation (file system → S3 → in-memory mock for tests) without touching the agent. -
Eval first, agent second. Both agents have an eval harness that runs against a small labelled fixture set on every commit. When I changed the code-review prompt last month, the eval caught a regression in the "false positive on
# TODOcomments" category before the change shipped. Without evals, you're just guessing whether your agent got better. -
Single model surface. All Claude calls go through
src/shared/llm.py. Want to try a different model? One line change. Want to add a retry on rate limit? One change. Don't sprinkleanthropic.Anthropic()across 12 files. -
Honest scope. These agents are good at enumeration (test categories, possible bugs, edge cases). They're average at judgment (is this enum exhaustive? is this bug a real one?). I curate. The agent's job is to bring me 30 candidate test cases in 90 seconds; my job is to keep the 15 that matter. Anyone who tells you their agent removes the human in the loop is selling something.
Eval results
Numbers from the internal version. Public version's eval set is much smaller; results here are illustrative.
| Agent | Metric | Internal | Public eval |
|---|---|---|---|
| code_review | Agreement with senior reviewer on real PRs | 78% | 71% on fixture set |
| code_review | False positive rate (review comment a human would reject) | 14% | 18% |
| test_gen | Coverage of edge cases vs. human-written tests | +37% | +28% on fixture |
| test_gen | Generated tests that compile and pass on first try | 91% | 84% |
What this is NOT
- Not a framework. No
pip install claude-pod. The code is meant to be read and forked. - Not a product. No web UI, no SaaS, no monitoring dashboard. That's intentional — this is the layer that should live inside your team's tooling, not as another service to deploy.
- Not infinite scope. Two agents, one domain (engineering automation), one model provider (Anthropic Claude). If you need more, fork.
Posts
- How a test-gen agent caught 15 edge cases my team would have missed (coming soon)
- Two years of running production AI agents inside a fintech team (coming soon)
License
MIT. Use freely. If you ship something with this and find it useful, I'd love to hear about it.
About
I'm Egor Tyakhti, a 20-year production engineer who's spent the last two years embedding LLMs into a 14-person fintech team's daily workflow. Currently exploring Applied AI / Forward Deployed Engineer roles. LinkedIn · Email