I keep building the same idea from different angles: autonomous agents doing consequential work, and the machinery that lets you trust the result without trusting the agent. Verifiable evidence, hash-chained provenance, deterministic adjudication, enforcement on actions. duelswarm is where it all converges.
flagship
duelswarm
A sandboxed cyber range where an autonomous red agent and an autonomous blue agent fight, both driven by language models, with every action checked by a deterministic authorization gate before it runs and every event written to a hash-chained log you can verify instead of trust. What it proves: you can let the agents be as capable and adaptive as you want and still hold a hard safety guarantee, because the thing that grants or denies an action has no judgment in it to fool. It is the instrument behind most of the writing on this site.
Read: Building duelswarm →adjudicating evidence
Noctelle
A production system for adjudicating cyber-insurance claim evidence. It takes messy, contested, real-world evidence and produces a defensible determination, with SHA-256 hash-chained provenance and byte-identical reproducible output, so the result can be audited rather than taken on faith. Built test-first, with no third-party runtime dependencies in the core engine. This is where the instinct behind duelswarm's gate and log came from: if you are going to let software make a consequential call, the call has to be reconstructable from the record, not believed.
autonomous offense
breachswarm
A multi-agent CTF solver built on the Claude Agent SDK. An orchestrator routes challenges to category specialists that run real tools in an isolated sandbox. It scored 70 out of 100 at pass@1 on InterCode-CTF, and I caught and fixed a benchmark leak, gold flags readable in the sandbox, before reporting that number. It proves autonomous multi-agent systems can do real offensive security on their own, and its orchestrator-and-specialists structure is the foundation duelswarm's red team grew out of. Public repo: link to come.
evidence for agentic commerce
Readback
An entry for the PayPal AI Hackathon. When an autonomous agent acts on your behalf and a transaction is later disputed, what is the evidence that you actually authorized it? Readback builds that delegation-evidence trail for dispute defense, with no server-side data storage, aimed at the claims-investigator workflow. Same theme as the rest: when agents take consequential actions, you need a verifiable record of what was authorized, not a story reconstructed after the fact.
Why these belong together
Offense that runs itself (breachswarm). Evidence a machine can defend (Noctelle, Readback). Enforcement that holds when the agent is adversarial (duelswarm). The common question under all of it is the one I care about most: how do you trust what an autonomous system did without trusting the system. The answer I keep arriving at is that you do not trust it. You verify it, and you enforce on its actions. That is the research.