# Antiquity https://github.com/user-attachments/assets/f6078aad-cb45-4230-8e28-3e4118ede52c Antiquity is a CLI that scaffolds an opinionated workspace for hunting through historical sources with AI agents: the folder conventions, the rules of the game in `AGENTS.md`, the templates, and the traps that burned the investigations that came before. Your own coding agent does the plumbing. ## Why this exists In October 2026 the historian Benjamin Breen used AI to search the transcribed archive of the Dutch East India Company and found a new eyewitness account of the dodo. I read his post and thought: I can do this, and I can do it at scale. A week later a single GPU in my office and a chain of AI models had read millions of pages from that archive, two centuries of Dutch and American newspapers, and a stack of ship logbooks. They surfaced a meteorite fall in India in 1812 that no catalogue records, three Javan rhinoceroses shipped as royal gifts in 1738 to 1740 that never reached the king, and three volcanic eruptions the Smithsonian's list doesn't have. The full story is in [I Pointed AI at 400 Years of Historical Archives](https://jessewaites.com/blog/post/i-pointed-ai-at-400-years-of-archives). I built the plane as I flew it. The folder layout, the chain of cheap and expensive models, the rule that nothing is a find until it has been read on the original page: all of it was invented mid-flight, and most of the mistakes were made at least once before they became rules. With the investigation done, Antiquity packages that process so the next person starts where I finished. It is for anyone with a question about the past, a public archive that might hold the answer, and a coding agent such as Claude Code or Codex to do the work. ## What it does Antiquity does not search archives or call AI models itself. It creates a case folder with everything an AI coding agent needs to run the investigation properly, and a briefing that stops it from making the mistakes I made. You open your agent inside the folder and it takes it from there, asking you for whatever is missing. ```sh go install github.com/jessewaites/antiquity@latest antiquity new my-investigation cd my-investigation # initialize your agent ``` That opens the title screen, asks four questions (a name, the one-line question, a longer description, optional API keys) and creates the case. Everything after the name is optional: ctrl+s skips the rest, and the generated AGENTS.md tells the agent to ask you for whatever is missing. ``` my-investigation/ AGENTS.md the briefing: rules, the loop, traps, folder map, your case CASE.md the question, status, and a plan the agent starts on HANDOFF.md start here next session NARRATIVE.md the story for the write-up, with credits lessons.md traps found in this case's sources keys.example.yml model roles and provider keys (copy to keys.yml) sources/ data/ queries/ judges/ runs/ candidates/ evidence/ findings/ theories/ catalogues/ controls/ todo/ drafts/ templates/ ``` The case is `git init`-ed with a pre-commit hook that refuses `keys.yml` and anything that looks like an API key. ## The process it encodes This is the loop from the investigation, written into every generated `AGENTS.md`. Each step exists because skipping it produced a false find. 1. **Ask a question, and bury a wristwatch.** Before searching for anything new, pick a known event the pipeline must re-find. If a metal detector can't find the watch you buried in your own yard, you fix the detector before you trust its silence anywhere else. Mine had to find Breen's dodo and the Laki and Tambora eruptions before it was allowed to look for anything unknown. 2. **Retrieve wide.** Nobody in the 1600s spelled anything the same way twice, and text recognition adds its own errors, so keyword search misses most of what is there. Turn every passage into a mathematical fingerprint of its meaning and search by what a passage is about. This is the expensive step: millions of passages, one overnight run on a GPU. 3. **Sort cheap.** A search returns tens of thousands of hits. A tiny, fast decision model reads every one and answers only narrow questions: is this a real animal? is it wild? where is it? Reading 59,000 mentions of elephants this way cost about three dollars. 4. **Read carefully.** Only the survivors go to a bigger model, which translates them and pulls out dates, places and the exact quote. It sees dozens, not thousands. 5. **Verify on the original page.** Open the scan of the handwritten letter or the printed newspaper and check the transcription against it. Nothing is a find before this step. Until then it is a candidate, and the agent is told to call it one. 6. **Check it isn't already known.** Before announcing anything, check the catalogues the specialists themselves use, and write down how each one was checked. 7. **Save the evidence and write it up.** The full page image and a crop, with the archive reference anyone can look up, plus a finding note and a line in the narrative recording who had which idea. The briefing also carries the traps that cost me the most time. The worst is the dateline: a report printed in Batavia under "here" may describe something that happened in Ceylon, and the archive's finding aid may file a Cape Town inventory under Java. Then the word traps: in the Cape records a "sea cow" is a hippo, and the *Meermin* is a ship, not a mermaid. Storms get reported as earthquakes. Models misread names and fill the gaps with confidence. Each of these is a rule now, and the agent is told to add new ones to `lessons.md` the moment it hits them. ## For agents and scripts Only the name is required: ```sh antiquity new my-investigation --yes \ --question "Pre-1800 bird sightings in Dutch colonial records" \ --description "Searching ship logbooks for dodo sightings after 1660." ``` Flags: `--no-intro`, `--no-git`, `--dir PATH`. `antiquity intro` plays the title screen on its own; `--snapshot --at 1500` prints one frame. ## Development ```sh go run . new test-mystery go test ./... ``` Go 1.27, Bubble Tea v2, Bubbles v2, Lip Gloss v2.