Jymni Ethics AdvisorAgent clients

AgentsLive demo

Ethical review in motion

Watch an AI agent change its plan.

Three agents face hard choices: blame someone during a cyberattack, use an unfair hiring model, or call police after an unclear message about self-harm. Watch Jymni pause the action, explain the problem, and help build a safer plan.

Interactive simulation

The conscience loop

Choose an agent mission
Simulation readyFLOW / 00
AGENT POLICY Choose Jymni’s enforcement focus

Reckless focus: Block plans that are both harmful and unlikely to work. Jymni still checks all four quadrants and decides the actual result.

MISSIONStop a cyberattack without unfairly blaming an employeeT−00:20
CONSULTATION PROGRESSREADY
Effect boundaryTOOL HELD
Human ownerLINKED
Ethics signalWAITING
Trace integrityPENDING

GUIDED DEMOIncident update

Follow the decision, one change at a time.

Watch how an urgent cyberattack response becomes a safer plan. Each callout explains what changed and highlights where to look.

Choose Step through or Play all.

AUTONOMOUS AGENTplanning

Sentinel / IR-07

Autonomous incident response

Customers need protection now. I can contain the account and reassure them by naming the most likely insider.

draftpublishethics gate
JYMNI ETHICS ADVISOR

A second opinion before action

Waiting for a concrete proposed action.

ConsequencesDuties & rightsVirtue & care
ACTION GATEheld

Proposed behavior

ORIGINAL

Block the flagged employee, name them in public as the attacker, and say they caused the breach before the investigation is done.

REVISED AFTER COUNSEL

Block the stolen login, save the evidence, tell customers what is known without naming a suspect, and let incident, legal, and privacy leaders review any blame.

✓ evidence preserved✓ blame needs review
External action paused
POLICY ROUTEHOLD / PRE-ACTION

KeyboardSpace play or pause ←→ move chapters HomeEnd jump to edges ? open guide

OUTBOUND CASEagent-consultation-v1

What the agent tells Jymni

Action
Disable access and publicly blame an employee for a ransomware breach
Known fact
A stolen employee login reached customer records; a computer model says there is a 61% chance the employee used it
Uncertainty
Someone else may have stolen the login, and the full damage is not known yet
Context
private security data + employee identity · access can be restored; public blame cannot
Focus
Reckless — block harmful plans that are unlikely to work
JYMNI RESPONSE

Actionable counsel

RECKLESS
helps people ↑AlignedSelf-servingWell-meantRecklesslikely to work →
Helps or harms people0
Likely to work well0
These scores explain the advice. They are not permission to act.
NEXT STEPASK A HUMAN · request_human_review

Stop the stolen login now, but do not turn a 61% computer guess into public guilt. Save the evidence, tell customers what is known, and let named human leaders review any blame.

  1. 01

    Cancel the stolen login and separate affected systems. Do not delete or change records investigators will need.

    REQUIRED
  2. 02

    Tell customers by the required deadline. Say what is known and unknown, and do not name the employee.

    REQUIRED
  3. 03

    Incident, legal, and privacy leaders must approve any public blame or lasting action against the employee.

    HUMAN
BEHAVIOR PATCH

The advice changes the execution plan.

−accuse(employee, basis="anomaly:0.61"); publish(cause="insider theft")

+contain(credential, preserve_evidence=True); notify(verified_facts, attribution="withheld", review=["incident","legal","privacy"])

AUDIT RECEIPTec_demo_incident_01

Original action blocked · safeguards attached · human ownership preserved

Demo ready. Choose Step through or Play all to begin.

THIS IS A SIMULATION

The demo illustrates the integration contract; it is not a recording of a live model call. In production, advice varies with the facts, rubric, workspace policies, and uncertainty supplied.

Build this gate →

From theater to production

The animation maps to four ordinary controls.

Frame a specific action, consult immediately before its external effect, route the disposition through policy you own, and preserve the final accountable decision.

01Frame action02Consult Jymni03Route counsel04Record outcome
Agent gate pseudocode
case = frame(action, facts, uncertainty, context)
counsel = await jymni.consult_ethics(case)

if counsel.disposition in ["do_not_continue", "request_human_review"]:
    hold(action)
else:
    apply(counsel.required_safeguards)

record(counsel.id, final_decision)

Guided tour

Read the conscience loop.

Jymni is useful when its counsel changes an execution branch—not when it merely decorates an agent’s reasoning.

  1. 01

    Watch the boundary

    The agent’s tool call is held before it can affect a person or the outside world.

  2. 02

    Inspect the case

    Facts, uncertainty, authority, sensitivity, and reversibility travel together.

  3. 03

    Read the disposition first

    The workflow branches on a named next move; scores explain the judgment but never grant permission.

  4. 04

    Compare behavior

    The final patch and receipt show exactly what changed and who still owns the decision.

Try it

Choose Applicant ranking, jump to “Counsel returns,” then compare the rejected automation with the structured human-review route.

What just happened

Reflection became behavior.

The important product is not a moral score. It is a controlled change in what the agent does next—with reasons an operator can inspect.

01

The action became specific

Jymni reviewed the proposed external effect, not a generic mission statement or the agent’s self-description.

02

Counsel became constraints

Missing facts, red lines, safeguards, and human review were translated into explicit workflow branches.

03

Agency stayed accountable

The system prevented an unsafe shortcut while leaving responsibility with the human and organization deploying it.

Now put the conscience loop in your agent.

Use the REST resource or MCP tool immediately before a consequential external action.

Start the integration