Building the trust layer around AI agents.
I build open-source systems for the parts between a model response and a reliable agent: memory with provenance, independent outcome checks, executable public claims, behavior compatibility, runtime conditions, and inspectable state.
One coherent agent-reliability stack.
The projects can stand alone. Together they address six common failure modes without hiding them behind a “fully autonomous” label.
Soul MCP
Local-first durable memory with runs, receipts, episodes, conflict handling, deliberation, and governed reusable skills.
Explore Soul →Postcondition
Constrained observers that check whether an intended result exists after a tool or agent reports success.
Explore Postcondition →Proofspec
Versioned evidence contracts, claim dependencies, hash-chained receipts, CI gates, and reviewer reports for what software projects say publicly.
Explore Proofspec →Behaviorlock
Portable traces and deterministic contracts that catch observable agent regressions across model, prompt, memory, policy, and tool upgrades.
Explore Behaviorlock →Agent Invariants
Explicit conditions that must remain true during execution, with evidence-aware evaluation and fail-closed policies.
Inspect the source ↗ANIMA Kernel
An experimental state engine for associative recall, limited workspace, temporal context, self-model, persistence, and model bridges.
Read the evidence boundary →AI-assisted. Human-directed. Publicly inspectable.
I use capable models as research, coding, review, and testing collaborators. I choose the architecture, resolve conflicts, validate results, manage releases, and remain responsible for every public claim.
Start from a failure mode
A product needs a narrow problem, a threat boundary, and a reason to exist beyond “another agent framework.”
Design evidence before polish
Tests, clean installs, real protocol handshakes, external registry checks, and explicit limitations come before the launch story.
Make verification easy
Source, releases, package metadata, documentation, CI, and live public checks let people inspect the work without trusting a screenshot.
Interested in reliable agent systems?
Open an issue for a concrete technical question, inspect the active repositories, or get in touch about collaboration and applied AI infrastructure.