if you let an AI agent run on its own, two things should scare you. it spends money you didn't approve, and it changes files you didn't check
most agent safety tools block a bad action and stop there. the harder part is proof: after the agent runs, can you verify what it changed and what it spent?
two tools, one stack. mythos-sentinel signs a spending limit the agent can't exceed and seals each session. mythos-router applies every file change under write discipline with a hash-verified receipt. runs locally
npm i -g mythos-router mythos-sentinel
https://t.co/xDtjvOx8s6
v1.18.1
mostly a stability and correctness pass across the main runtime
changes:
• orchestrator retry logic now uses authoritative status codes, which fixes accidental retry loops
• FTS5 search is now properly tokenized, so inputs like c++ or broken quotes no longer silently return empty results
• SWD execution now fully reports aborted batches, and rollback behavior is deterministic instead of sometimes missing history
• rollback system tightened, it only removes what the run actually created, prevents double rollbacks, and won’t touch existing structure anymore
• repository wide LF enforcement to kill cross-platform hash drift
Also starting the next phase: mythos-sentinel
When the Mythos leaks appeared in March, I started building around them, the missing layer.
Mythos Router enforces strict write discipline and validates state, giving powerful models guarantees they never had. As models become more capable, guarantees should not be optional.
With Mythos going public soon, I’ve been building around this for months.
your coding agent has lied to you. you just didn't check.
it emitted a clean tool call, said the file was written, and you believed it because the JSON was well formed. but a schema validates the request, not whether the write actually happened.
everyone's trying to make agents more capable, which is good, but almost nobody's making them prove what they actually did. sometimes a worse model you can easily verify could beat a genius you have to take at its word.
right now that's a gap. we're making verification the default, not the exception. what's next:
• diff based edits
• receipts you can replay to undo a verified change
• verification that plugs into the tools you already use
@elonmusk This guy is going to take over King’s Landing one day. Game of thrones is on but people think this guy is a nobody. They will find out when he sits as the rightful heir to the iron throne.
Against the odds 🤣
Over the weekend we launched Path of Exile 2: Return of the Ancients. We're thrilled to see so many of you jumping in and exploring all of the new content. We wanted to take a moment to share an update on how the launch has gone so far and highlight the issues we're currently prioritising https://t.co/tnAkzYpnFA
v1.16.0
a lot of agentic focus is on payments and onchain actions right now.
this release focuses on a different side, what happens when agents need to touch real codebases?
mythos router can now take proposed file changes from an external agent, validate them, keep them inside task boundaries, run checks in a temporary repository copy, and record the outcome before applying.
agent execution needs more than prompts, it needs guardrails too.
700+ weekly npm downloads!
pushing an update tomorrow, moving further toward structured agent runs:
• action schema and validation
• task boundaries
• run outcomes
• policy suggestions
nexus updates are coming soon too, a few new additions