Four systems that read a repo, draft the mail, pull the filings and publish the report. No prompt box, no operator, no demo mode — a schedule, a toolbelt, and a ledger that records what every run cost.
Every system on this page walks the same five stages. Only the source and the toolbelt change. Pick one.
Give it a local path or a GitHub URL. It clones, indexes and reads the codebase, then writes tickets an engineer can actually act on — title, description, acceptance criteria, priority. Served over REST. No database anywhere in it.
Multi-agent office automation on CrewAI and GPT-4o. Four crews draft email, summarise documents and run US equity analysis on a schedule. Every run is costed into SQLite and surfaced in a Streamlit dashboard, so the bill is never a surprise.
Collects SEC filings — 10-K, 10-Q, 13-F, 6-K — for 30+ US equities, structures them into fundamental, insider and technical reports with Claude, then rebuilds and publishes the whole site daily from GitHub Actions. Nobody presses build.
21 structured prompt frameworks that turn any LLM into an equity analyst — Piotroski scoring, DCF, 13-F tracking, earnings-call sentiment. Slash commands inside Claude Code, plain text everywhere else. Nothing to install, nothing to run.
Cut the same four systems the other way and you get three layers. A run enters at the top and does not stop until it has left a row at the bottom.
One entry point, four crews that share nothing, one ledger. That is the whole trick: a failed crew is a failed crew, not a failed system.
┌────────────────────┐
│ main.py │
│ cli entrypoint │
└──────────┬─────────┘
│
┌──────────▼─────────┐
│ APScheduler │
│reads schedule.yaml │
└──────────┬─────────┘
│
┌────────────────┬────────┴───────┬────────────────┐
▼ ▼ ▼ ▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ TicketCrew │ │ EmailCrew │ │ DocSummary │ │ StockCrew │
│ repo index │ │ gmail smtp │ │ pdf / docx │ │ yfinance │
└──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘
│ │ │ │
└────────────────┴────────┬───────┴────────────────┘
│
┌──────────▼─────────┐
│ OpenAI GPT-4o │
└──────────┬─────────┘
│
┌────────────▼───────────┐
│ runs.db (SQLite) │
│ tokens · cost · status │
└────────────────────────┘
A job that fires every weekday morning sends the same system prompt, the same toolbelt and the same output schema every single time. Attention already did that arithmetic yesterday. Keeping the prefix warm is the difference between paying for the preamble 250 times a year and paying for it once.
finance_data answers out of filings, not out of the model. Embedding search gets you close — close is a synonym, a stale press release, a paragraph that mentions the word. The reranker is where the top-k stops being trivia and starts being evidence.
An agent is four moves on repeat, and the fourth is the one people skip. Anything can observe, plan and act; a system you can leave alone overnight is one that checks its own work and then decides the run is over.
The verifier passes and the artefact exists on disk. Nothing else counts as done.
Tokens and dollars are decremented per step. At zero the run stops and says so.
One retry with backoff. A second failure is a fact about the world, not a prompt problem.
Twelve. A loop with no ceiling is not an agent, it is a bill.
Swapping one frontier model for another is an afternoon. The year goes on everything around it: what is allowed into the window, what gets thrown away, what happens on the second failure, and what is written down before anybody reads the answer.
Every step decrements a token and dollar allowance. The ledger is written before the output is read.
A retried send must not send twice. Side effects carry a key, not a hope.
An unparseable answer is a failed run, not a shipped one. Schema first, then the write.
Inputs, prompt hash and output land in runs.db, so yesterday’s eight o’clock can be re-run today.
That is the part I enjoy — the agent being right when nobody is checking. Every repo above is open. Mail me if one of them is useful to you.