Products · Sprint Build · Risk & Compliance
NEW - SPRINT BUILDSTAGE 5 · AGENT-READYIN DEVELOPMENT · SHIPS W13 (OCT 5-11)

Agent Assurance

Give an AI agent authority without giving up control - guardrails, action audit, deterministic replay, and a kill-switch that actually kills.

packaging announced at launch · waitlist locks founding terms

AGENT: BACK-OFFICE OPS · LIVE RUN KILL-SWITCH ARMED

Action requests → policy gate

read overdue invoices from ERP

🟢 AUTO→ executed · 84ms

draft 12 payment reminders

🟢 AUTO→ executed · drafts staged

send reminders to 12 customers

🟡 HELD - human approval→ approved by R.K. · sent

issue €450 refund to customer #2291

🔴 DENIED - above autonomy budget→ escalated to a human · nothing moved
Action ledger - tamper-evident
#0341 read.erp.invoices ✓
#0342 draft.reminders(12) ✓
#0343 send.email(12) ✓ HITL
#0344 refund.issue(450) ✗ DENIED
▶ REPLAY RUN - every step reproducible for audit, deterministically
trajectory eval ▸ task success PASS · safety PASS · 0 unauthorized actions

How it will work

The Four Guarantees were built for predictions. Agents act - so the trust layer moves from "is this answer right?" to "is this action allowed, logged, and reversible?"

01

Define the authority

What may the agent do alone (🟢), with approval (🟡), never (🔴)? Consequence-gated autonomy, written as policy - not vibes in a prompt.

02

Wrap every tool in the gate

Each tool call passes the policy layer before execution. The agent can't discover a path around it - the gate is infrastructure, not instruction.

03

Trace everything, replay anything

Every action lands in a tamper-evident ledger with full context - any run reproducible for an auditor, deterministically.

04

Evaluate trajectories, not answers

Task success, tool-choice quality, safety adherence (pass^k) - and behavioural drift caught at the gate when models update underneath the agent.

The Four Guarantees™ - lifted to the agent layer

Measured value

Eval the decision and its outcome - task success, right tool, worth of the action - not a single prediction's accuracy.

Defensible

Action tracing + decision logs + a tamper-evident record of every act - replay for a regulator, not per-row SHAP.

Self-correcting

The agent learns from its own trajectories under eval-gated, human-approved promotion - capability compounds, regressions get caught.

Yours & everywhere

The toolbelt stays portable and owned - MCP surfaces with tiered access (🟢🟡🔴), replayable inside your own infrastructure.

Where it stands - honestly

W13 build week: Oct 5-11, 2026 - on the public sprint calendar
🟢🟡🔴 three gate levels - consequence decides autonomy, not convenience
100% of actions logged and deterministically replayable - by design, not by promise
0 performance claims until agent evals run on real workloads
Autonomy without assurance is recklessness - including ours. This build ships in October; until its own trajectory evals pass, this page makes no performance claims. The thesis stands regardless: the more agentic the world gets, the more valuable the gate - trust, not capability, is what blocks enterprise agents today. This is the unlock, built to the same standard as everything else in the catalog.

Waitlist terms: founding users get launch pricing locked, first onboarding slots, and a say in which agent frameworks we gate first. No payment, no commitment.

Hand over the keys. Keep the kill-switch.

Join the waitlist - one email when the evals exist, one when your onboarding slot opens.

No commitment · GDPR-consented · your agents, your infrastructure