Simam Agent Harness

What needed solving

AI agents now call real tools: they push code, run shell commands, query databases and read files. Most teams have no single place that decides which of those calls an agent may make, stops a dangerous one before it runs, or keeps a record they could show an auditor afterwards.

How we're building it

Every agent tool call passes through one pipeline: an emergency kill switch, a per-agent tool allow-list, policy rules, governed operations, budget limits and, where needed, human approval. Governed operations are dangerous actions such as force-push, DROP TABLE, rm -rf and reading secrets, each mapped to the OWASP LLM Top 10, OWASP Agentic and MITRE ATLAS. Every decision, whether allowed, held for approval or blocked, is written to a tamper-evident evidence ledger with downloadable receipts. An MCP proxy brings any MCP server under the same rules, alongside local posture scanning, compliance-style reports and role-based console access.

What we learned

Built test-first with AI-assisted engineering: 369 backend tests and about 90 frontend tests. Governance only earns trust when every decision can be replayed: the ledger and its receipts matter as much as the block itself.

What we are testing next

Can the harness govern a coding agent live, checking every Bash, Write and Read that Claude Code attempts before it runs, without slowing the developer down?

Where it is

1. Sep 2026
   Gateway pipeline, governed operations and evidence ledger built test-first
2. Oct 2026
   Live beta with an invite-only console, MCP proxy and public demo instance
3. Next
   Govern Claude Code itself live, checking each tool call before it runs