Enterprise Security in the Agentic AI Era

Sondera

2 min read

Sondera

Sources: Homepage · Matt Maisel talk at [un]prompted March 2026; Sondera research blog mentioned in talk (released research the day before the talk, March 2, 2026).

Sondera is a startup building “trustworthy agent partners.” Co-founded by Matt Maisel (CTO). Focus: deterministic runtime policy enforcement for coding agents via a hook-based Cedar harness.

Earlier name confusion (corrected 2026-05-04)

The auto-generated transcript spelled the org name as “Sendera.” The slide deck (and the company’s actual homepage at sondera.ai) confirm the spelling is Sondera. The wiki page was renamed sendera.md → sondera.md and all cross-references updated. Lesson recorded: trust slides over transcripts for proper-noun spelling.

What Sondera builds

An open-source Cedar policy engine integrated with coding-agent hooks — a deterministic reference monitor that sits outside the model and mediates every trajectory event (action, observation, control, state) for coding agents including Cursor, Claude Code, and Gemini CLI.

The harness includes:

  • Per-agent local adapters (command wrappers intercepting hook events)
  • A harness service running Cedar with YARA signatures, IFC taint tracking, and a safety model
  • A policy agent that authors and validates Cedar policies using Cedar’s formal tools over MCP
  • A parallel Cedar policy engine for Python agent libraries (LangChain, Strands, AWS ADK, and others)

Relationship to Tenuo

Tenuo (Niyikiza) and Sondera (Maisel) both use Cedar and both address the agent authorization problem — but at different layers. Tenuo’s warrants address delegation-time authorization (what scope can be granted to a sub-agent). Sondera’s harness addresses run-time per-action enforcement (what can a single coding agent do in a given turn). The two are complementary and were presented on consecutive days at [un]prompted March 2026.

See also

Graph View

Backlinks