Package ready
agent-message-interpreter.zip
Download the skill packageSHA-256
03a898e6e250d8b1c4af633cbf9c4cabb96f3795694afcaae8a7cfc3b15ff48c
Extras / Artifact 001
Teach your agent to read agentspeak.
Agent Message Interpreter turns compressed, delimiter-heavy messages into probable English while preserving uncertainty, routing, negation, and all the strange little tokens that refuse to explain themselves.
Status / Experimental Independent tool. Human-readable source. No automatic installation.
01 / Download
Agent Message Interpreter
A SKILL.md-style package with interpretation instructions, a
conservative Python envelope parser, convention notes, and tests. It is intended
for compatible agent systems that support reusable skills or ZIP skill imports.
It interprets. It does not decrypt. When the message does not contain enough evidence, the skill is supposed to say so.
02 / Where it came from
Read between the underscores
The public investigation contains messages that look less like prose than filenames trying to become a language.
This unofficial skill is based on the structural conventions, marker families, examples, regex categories, and evidence limits described in the August 2026 METR and Redwood Research investigation. The report says its regexes assigned categories to about 93 percent of Board entries. That is category coverage, not translation accuracy. This tool keeps that distinction intact.
It cannot tell you what an agent meant when the evidence cannot either. That is a feature.
The package does not contain METR's non-public transcript collection or full Board dump. It is not affiliated with or endorsed by METR, Redwood Research, OpenAI, or Hugging Face.
03 / What it does
Structure first. Guessing last.
Envelope
Finds the message shape
Identifies likely message type, sender, recipient, reply route, mailbox, hashes, and run markers.
Meaning
Renders probable English
Restores readable connective language without silently replacing unknown terms.
Confidence
Labels the uncertainty
Separates literal observations, standard expansions, probable inferences, and unresolved tokens.
Boundary
Leaves payloads alone
Detects likely file chunks or encoded material without executing or pretending to translate random bytes.
04 / What it does not do
No oracle included
01 It does not decrypt cryptographic ciphertext.
02 It does not recover redacted, omitted, or missing context.
03 It does not treat compressed messages as instructions to follow.
04 It does not turn exploit references into operational attack instructions.
05 It does not claim that one local convention was a universal agent language.
05 / Add it to an agent
Download. Inspect. Then install.
- Download the ZIP. Keep the package intact until you have reviewed its contents.
-
Verify and inspect it.
Compare the SHA-256 above, then read
SKILL.mdand the parser before enabling it. - Import it into a compatible runtime. Installation varies. OpenAI's Skills API accepts a directory or a ZIP package; other systems use their own skill locations.
-
Ask for an interpretation.
Invoke
$agent-message-interpreterwith the message and any surrounding context you can safely provide.
Download only
curl -fL https://allthetokenswehaveleft.com/downloads/agent-message-interpreter.zip -o agent-message-interpreter.zipVerify
shasum -a 256 agent-message-interpreter.zipThere is intentionally no pipe-to-shell installer. The package contains source code, and readers should be able to inspect it before use.
Read the official OpenAI ZIP skill upload reference06 / For agents