Foundation models provide probabilistic reasoning. The Harness provides the deterministic runtime substrate: execution sandboxes, verification sensors, and recovery loops: Agent = Model + Harness
We haven't launched yet. But leave your email to get notified about exciting things that we will launch.
No spam. Technical updates and release announcements only.
Production harnesses are formal runtime systems. Built on the ETCLOVG architecture, a harness governs execution, tooling, context, lifecycle, observability, verification, and policies.
Hermetic Firecracker microVMs and gVisor sandboxes with non-interactive process supervisors (PAGER=cat) and Copy-on-Write git worktrees.
Model Context Protocol (MCP) and Agent-to-Agent (A2A) schema enforcement to prevent hallucinated tool calls and malformed parameters.
Overcoming the "Dumb Zone" through Agentic Context Management (ACM), Tree-sitter AST symbol maps, and progressive context disclosure.
Tessl Inner, Outer, and Meta loops combined with the Hashimoto Protocol (Mistake → Permanent Control) and state-machine checkpointing.
Sub-5-second Test Impact Analysis (TIA), compiler gates, ArchUnit structural boundaries, and Princeton SWE-agent ACI diagnostic formatting.
Allow/Ask/Deny policies, code churn auditing, auditable Episode Packages, and multi-agent coordination.
14 comprehensive research volumes with 32 academic and industry citations.
Have questions about harness engineering or interested in collaborating? Connect directly on LinkedIn.
Slava Razbash on LinkedIn