Did your AI actually read your files?

LLM APIs are stateless — every turn re-sends the whole conversation. Paste a 40 KB report into a subagent and you pay for it on every subsequent turn of that agent's life. Add a second subagent and the report lives in two context windows; add a fact-checker, three. Multi-agent systems run at ~15× the tokens of a chat, and roughly 37% of their failures trace to agents acting on divergent copies of what should have been the same information.

A raw file path is the right instinct — 30 bytes is the right size for a handoff. But a path has no attribution, no adaptation, no lifecycle, no telemetry, and no reach. waggle keeps the reference small and makes it answer back.

Source code is where the lens shines. --snapshot runs tree-sitter at mint and stores a symbol outline beside the bytes. The consumer orients before it greps, reads a definition by name, and you can declare — and prove — what a reviewer had to reach:

waggle mint  --target "file://…/contract.rs" --snapshot \
      --require symbol:evaluate     # a signed consumption contract
waggle read     --token 9u6KEr6F --symbol evaluate   # the exact definition
waggle coverage --token 9u6KEr6F              # { met: true } ← it was reached

A PDF reads like text. The substrate extracts a PDF's or an HTML page's text layer at mint — deterministically, and it records that it did — so read, search, and contracts work over the document itself. The extraction travels with the token, not as a loose file the next agent has to find. What it will not do is guess: audio, video, and scans carry no text layer, so waggle serves the bytes and tells the consumer to read them with its own model — because that model is better at it than any converter we could ship, and a receipt must never vouch for a transcript nobody can reproduce. In our runs a PDF answered 100% at 750 bytes, where pasting the whole file took 23,651.

Then wire it into your harness — all three harnesses land on the same daemon and the same tokens. What a Claude Code session mints, a Codex session resolves.

claude mcp add waggle -- waggle serve --stdio   # same line in Codex & Cursor
waggle init                                     # installs the agent stub

You don’t start a server. waggle serve --stdio is the MCP server; your harness launches it, and it auto-starts a shared background daemon the first time any harness talks to it. Nothing to run, nothing to keep alive. Restart your harness so it picks up the new tools, then confirm:

waggle daemon status   # → running · uptime · connections · db size

The store lives at ~/.waggle/waggle.db (SQLite). To run the daemon in the foreground for debugging instead of letting it auto-start: waggle serve --daemon.

npx wrangler deploy      # a Durable Object per tenant
waggle edge push         # records + snapshots replicate; the FILES never leave
waggle edge status       # { "health": "ok", "tools": 9 }

An agent's loop — mint, hand off, resolve, interrogate, report — is byte-for-byte the same whether the other end is in this process, on another machine, or on another continent.

This is stigmergy — coordination through durable marks rather than direct messages — and it is the same discipline distributed systems spent forty years converging on: tuple spaces, named-data networking, capabilities, leases, content addressing, the log-as-truth. The essay traces the whole lineage.