EU AI Act Article 14 enforceable 2 August 2026 · · Penalties to €35M or 6% of global turnover

Live supervision for agent runs

Your agent has been running for six hours.
It drifted at minute forty.

Every human-in-the-loop tool is an approval queue: one action, one person, one click. compass supervises the whole run — your team joins mid-flight, steers it together, and leaves proof a human was in control.

run/4471 · settlement-reconciliation-agent 00:00:00
M
R
A
Live · agent is working

The gap

An approval queue can't supervise a process.

The industry solved oversight for a world where an agent did one thing and asked permission. Agents now run for hours and take hundreds of actions. Those are different problems and only one of them has a product.

What exists today

The approval queue

Oracle Integration, Cloudflare Agents, LangGraph interrupts, HITL SDKs.

  • One designated reviewer per gate
  • Pauses on a single discrete action
  • Approve or reject — no way to redirect
  • No shared view of the run in progress
  • Reviewer sees the action, not the six hours behind it
  • Catches a bad action; misses a bad trajectory

What compass does

Live supervision

The run is the unit. The room is the interface.

  • Anyone on the team can watch a run in progress
  • Join mid-flight with a narrative handoff, not a transcript
  • Steer, redirect, and hand off between operators
  • An explicit rule for two people steering at once
  • Halt from anywhere, by anyone with authority
  • Every intervention attributed and retained
~70%
of real office tasks are failed by autonomous agents
CARNEGIE MELLON
20–30%
of long agent runs require human intervention
PRACTITIONER TESTING, 2026
4%
of organisations have fully operationalised AI
SRE ADOPTION SURVEY
€35M
or 6% of global turnover — Article 14 exposure
EU AI ACT, FROM 2 AUG 2026

The hard part

Joining at hour three.

A database stores facts. A handoff tells a story. Dropping someone into 190 tool calls isn't supervision — it's homework. Compare what a teammate gets today with what compass hands them.

run/4471 · joined at 03:11:42 step 148 / ~190

State snapshot

Typed, current, small. What the run believes to be true now — not the path it took to get here.

Narrative context

Three sentences on why the state looks like this. The part a transcript can never give you, because intent isn't in the log.

Decisions & dead ends

What was decided, deferred, and abandoned. Rediscovering a dead end is the most expensive thing a joiner can do.

Priority queue

What to look at first, second, third — so joining costs thirty seconds instead of thirty minutes.

Concurrent control

Two people. One run. Who wins?

Every shared-agent product has to answer this and none of them do. Documents merge; processes with side effects don't. Send both instructions and watch the resolution.

MAYA · staff engineer

"Stop retrying. Fail closed and alert on the first 5xx."

RAVI · payment ops

"Keep retrying but cap at three, then queue to the DLQ."

Resolution

Send one or both instructions to see how compass resolves them.

Article 14 & 26

Proof a human was actually in control.

From 2 August 2026, override capability, halt capability and a retained intervention record stop being good practice and become a legal obligation. Everything you did on this page was recorded.

TimeEventActor
No interventions yet — use "Take the wheel" or send a steering instruction above.
Article 14 — human oversight

Oversight persons must be able to interpret the system's output, override or disregard it, halt it via a stop button or comparable procedure, and remain aware of automation bias. compass is that surface.

Article 26 — deployer accountability

You are accountable for the AI you deploy regardless of who built it, and must retain logs of override and halt events for a minimum of six months. compass exports them.

Beyond the EU

The same architecture satisfies the NIST AI RMF GOVERN function and ISO/IEC 42001 human-review requirements. The vocabulary differs by regulator; the evidence trail doesn't.

Private beta

Nobody should have to babysit their own agent.

We're onboarding teams running agents that take irreversible actions — payments, finance ops, production systems.