I'm a Salesforce architect who builds agent systems for my own work. These repos are the parts I ended up reusing: guardrails that stop an agent from acting without a human, evals that check its output, and the plumbing to keep many sessions straight. Each one runs from a single command with sample data, and each README has a Results section with counts from my own use.
More about them: https://made-by-asking.pages.dev/showcase/
- prod-lockdown-guards: writes to production blocked until a human clicks. 4,509 tool calls checked in three days, 11 denied.
- consent-tokens: typed approval for payments and phone access. 30,323 checks across the two guards, 66 denials.
- studio-credit-guard: a paid generation runs only after a typed approval. 32 requests filed, 12 never approved, and none of those spent a credit.
- qa-loop: build, then an adversary tries to break it. Of 62 runs that passed, 43 passed only after a failing first round.
- dominio-router: injects only the area files a prompt touches. Context was added on 92 of 902 prompts.
- thread-notes: one note per work item with fixed fields and a checker. 65 notes, 64 done.
- whatsapp-memory-mcp: local search over exported chats as an MCP server. 18 tests; it has no usage numbers yet.
- session-board: every Claude Code session as a card. 1,835 session transcripts read over two months.
- usage-brake: a daily allowance with alarm levels for plan usage. 58 alarm events over 40 days.
- release-pollers: one poller per awaited change, one email, then it stops. 56 runs, 1 email sent.
- live-call-coach: cards for an interview or meeting as it happens. 19 tests; the latencies in the demo are simulated, so there are no field numbers.