What we changed in xNAUT, why we changed it, and where the idea came from. We research widely, listen to what is not working, and let the better approach win.
A cloud VM is not the problem. Every environment that starts empty each run erases the state a whole class of faults lives in, and for an app whose agents write their own history, that state is the product. Part 2 of 3.
Read →Two fixes, both correct, both tested. Together they built a retry loop that could never stop, because the brake counted evidence the improved code no longer produced. Part 1 of 3.
Read →Twenty-five years ago EAI engineers built typed messages, durable queues, guard-rails and rollback for endpoints that could not think. The endpoints just learned to think, and nothing else changed. Old skills are not dead; they were waiting.
Read →An engineer at Securosys read one sentence on this site and asked two questions. One was a word. The other was a hole: every attestation receipt was signed on its own, so deleting one left every remaining signature valid. What we changed, and the three details that would have quietly ruined it.
Read →Agents act on their own, so the record of what they did is the only control surface you keep. A hash chain proves nobody edited a row; it cannot prove nobody rewrote the file. Why v1.17.0 puts a Swiss HSM under the log, and what a signature honestly does not prove.
Read →A message to an agent is a question again, a build asks where before it touches anything, and 148 MCP servers are one click from any agent.
Read →What shipped in 1.14.0: build slices held to what they promised to produce, verification where the work ran, and a release that offered to update itself to the version it already was.
Read →NautLoop and GitVM: sovereign sandboxes for agents, pinned to a jurisdiction, with a control plane that holds none of your data.
Read →Prompt-to-app tools make you discover the interface by reading the output. Draw the screens first, argue while they are cheap, then build.
Read →Eight upgrades are landing in xNAUT's build stage in v1.13. Almost none of them started as our idea. Here is what research and community feedback changed, and what is coming next.
Read →Most "AI builds your app" tools take a prompt and start writing files. NautFlow writes the documents first, has a second model attack them, gets your approval on a design, and only then lets several agents build in parallel, judged by an acceptance test written before the build began.
Read →I got fed up with my tools, so I built my own terminal, then my own agent flow, then the part that proves what the agents actually did.
Read →