Structural safety for autonomous AI
THRESHOLD SIGNALWORKS
Safety infrastructure for AI agents
Your AI agent has access to your email, your files, your calendar. We build the structural guarantees that stop it doing something irreversible without your informed consent.
Threshold Signalworks builds safety infrastructure for autonomous AI systems. Not prompts that ask the model to be careful. Architecture that makes catastrophic mistakes mechanically harder. Guardrails that survive context compaction (when long conversations are summarised and instructions get dropped). Audit trails that prove what happened.
For people building agent workflows that can email, write, delete, and deploy.
Products
Prevent irreversible actions. Classifies every agent action by risk, requires structured human approval before anything destructive, and logs everything to a tamper-evident audit trail. Constraints live on disk, not in context, so they survive when conversations are compacted. Available now for OpenClaw, Claude Code, and any MCP-compatible agent.
Measure drift and regressions. Tracks whether your model still behaves the way you expect across version updates, prompt changes, and workflow modifications. Reproducible evaluation runs with full provenance chains, so anyone can re-run and verify.
Score uncertainty before acting. Detects when a model is guessing rather than reasoning, flags premature convergence, and provides calibrated confidence signals that feed into Keel's risk assessment. Low confidence on an irreversible action means stronger approval required.
Persistent safety for teams. Syncs policies and audit logs across multiple agents, provides a web dashboard for reviewing agent activity, and produces compliance-ready audit exports. EU-hosted, GDPR-native.
Research
Threshold Systems is the research arm of Threshold Signalworks. We study how instability enters during inference, tool use, and autonomous workflow execution, and we build measurement and intervention tools grounded in that understanding.
Current work spans AI evaluation protocols, cognitive architecture under constraint, and human decision-making in high-uncertainty environments. Publications and artefact packs are released through threshold.systems.
Artefacts
Public artefact packs (evaluation runs, reports, provenance chains) will appear here as they are released.