Resources
Read the work
The white paper behind Tapestry, what we have written about autonomous systems, verification, security, and the cost of supervision, and the code we have put in the open.
Documents
-
Tapestry: A Trustworthy Autonomous System
The full explanation of Tapestry and its first persona, the Python Developer. Why supervision is the cost of agentic AI, how the operating model removes it, the five core elements and what each one carries, the boundary every request and release crosses, and the evidence standard the work is held to.
Read the white paper (PDF)
Articles
-
An Attack Needs a Path
Hostile content cannot cause a consequential action by itself. It needs a reachable input, a susceptible interpretation step, a state transition, and permission to act. Agent security starts by tracing that path through the actual architecture and breaking it before untrusted language can acquire authority.
Read the article -
What the Human Was Actually Doing
Approval was never the whole job. A reviewer holds the requirement steady while the model drifts, reads a failed test and decides whether it matters, and chooses what happens next. Take the person out and that work moves into the system, where most of it turns out to be scheduling, lookup, and test execution, and belongs on a CPU rather than in a model.
Read the article -
Attention Is the Scarce Resource
OpenAI is testing a Codex mode that runs until it is put to sleep, and it ships as the most demanding setting in the reasoning-effort menu. Persistence bought with more inference does not reduce what agentic work costs, because the constraint was never tokens. It is the attention it takes to keep the work moving.
Read the article -
Play Is How Systems May Learn What Can’t Be Found
Some facts about a deployed system were never written down. They appear only when an agent tries the unexpected input, combines the components, or watches the sandbox fail. Play turns interaction into first-hand, system-specific observations that a corpus cannot supply.
Read the article -
Proofs Compound. Pattern Matching Doesn’t.
Scaling makes pattern matching better. A system that keeps validated proofs builds something different: a queryable, citable corpus of reasoning that survives model replacement and lets new work extend what has already been established.
Read the article
Open Source
Available now
-
Engram
A lightweight, keyword-indexed retrieval layer that returns a held answer before the expensive call ever runs. Legible by design, it learns from use and stays small. Clone it and see for yourself.
View on GitHub