• STOAT: local QEMU VMs for people and agents. Disposable Alpine sessions and persistent Linux guests from plain TOML, driven through a TUI, a JSON CLI or MCP tools. One Go binary.
• MIDWIRE: closed-loop agency. Confirms that a tool call's side effect landed before the agent takes its next step.
• DOCKET: an append-only decision ledger for coding agents. When a conversation resumes or gets compacted, it briefs the agent on what was decided, why, and what is still open.
• OPENCODE-NATIVE-SWARMS: permission-scoped background-agent workflows for OpenCode. It can research, review and run narrow checks, and has no path to edits, pushes or external writes.
• SKILLOPT: trains reusable natural-language skills for frozen agents through trajectory-driven edits and validation-gated updates.
• MONEY-MESH (private): a leaderless mesh of self-replicating earning agents under an immutable core. A Cedar policy engine sits at the enforcement point, revenue counts only from ground truth, and the spend cap is enforced by arithmetic.

• PALPATINE: a "strategic" advisor for Claude Code and Codex. It names the actual problem and the actions to take, in 50 words.
• CURT: editing guidance and a heuristic prose linter that cuts AI slop and keeps the author's voice.
• READABLE-RESPONSES: a readability hook and a de-slopify skill for Claude Code and Codex CLI, because Opus 5 yaps too much.
• POLISHED-ARTIFACTS: one design language for Claude artifacts, in Carbon and Apple looks, with a lint for rendered pages.

Memory that separates what an agent observed from what it concluded and what it made up.

• ENGRAMMIC: epistemic memory as a graph of claims, evidence and provenance.
• VEIL: drop-in local-first agent memory with sqlite-vec embeddings, FSRS decay and deterministic eviction.
• LEAP/CITE: a layered epistemic protocol with stratified types and coherence enforced at write time.

• ØCLOAK (private): at-cost RF and WiFi-sensing hardware plus a peer-to-peer threat-intel network. Surveillance got cheap; defense didn't.
• TAPESTRY: sovereign frontier models. Nations train one shared model and own their derivatives.
• GOOB: a desktop pet that roams your screen.

• Can an agent get a disposable machine as easily as a shell? (stoat)
• Can decisions outlive the session that made them? (docket)
• Can skills be learned in text space, without touching weights? (SkillOpt)
• Can stratified, warrant-backed memory run locally? (LeAP, shipped as Veil)
• Can an agent confirm its side effects before it continues? (midwire, in progress)
• Generation got cheap and verification didn't. What closes that gap? (essay)
• Can autonomy be bounded by a policy engine and arithmetic, rather than by the prompt?
• What should a robot remember? Object states, failures and maps, retrieved under partial observability without surfacing stale entries.
• World action models: predicting the consequences of an action in latent space instead of rendering future frames.
• Closing the loop on hardware: a VLA policy that checks its action landed through sensors, not through its own prediction.
• Safety bounds a VLA policy cannot override: runtime envelopes enforced outside the model.
• Cross-embodiment transfer: data from one robot's kinematics and sensors that still teaches another.