A runtime for agents that act. Palo Alto · May 13, 2026

Trust, shipped.

Agents worthy of the trust you place in them.

I. The problem.

Autonomy without accountability is just faster chaos.

Frameworks got us to deploy. Dashboards tell us what already broke. Nothing sits at the moment that matters — the instant between what the agent decided and what the world is about to remember.

Agents don't crash. They commit.

Salus sits on the action path. Every tool call, checked against the evidence behind it. Anything ungrounded is refused, with feedback the model can read and act on. No retrospect. No postmortem. No tax.

II. A moment

Eight milliseconds before the wrong thing happens.

BLOCKED
PREFLIGHT · 8MS

issue_refund(order="#4821", amount=420.00)

because, lookup_order returned $42.00. The model confused dollars with cents.

↳ issue_refund(…, amount=42.00) ✓ committed

BLOCKED
PREFLIGHT · 11MS

send_email(to="all_customers", subject="your appt is confirmed")

because, the agent was asked to confirm one booking. The recipient set expanded to 14,203 addresses. No evidence grounded that audience.

↳ send_email(to="jwhitley@acme.co", …) ✓ committed

III. What we believe

Three opinions, held firmly.

Monitoring is mourning.

Watching your agent fail faster is not safety. Stopping the action is. The dashboard is the obituary. Salus is the intervention.

ii.

Evaluations are not evidence.

A passing test set is not a promise to your customers. What survives staging is not what survives production. Trust is earned on the hot path, one tool call at a time.

iii.

Not every action deserves to run.

Every runtime is permissive by default. Salus is the first one that refuses. Every commit must cite the evidence behind it. That is most bugs.

IV. Get in touch

If your agents touch anything real,
we should talk.

Thirty minutes. Bring an agent. Watch Salus replay yesterday's sessions and show you what would have been allowed, clarified, repaired, rewritten, and escalated.

Book a demo →

founders@usesalus.ai

❦