human-only step
Agents stop where responsibility starts
Some steps require judgment, context, validation, negotiation, presence, or accountability that should not be faked by automation.
agentic HITL orchestrationWhatsApp / text / video
When agents hit a human-only step, prox.city handles the human handoff.
prox.city is a human-in-the-loop orchestration layer for agentic workflows. When an agent reaches a step that needs judgment, context, validation, responsibility, trust, negotiation, or human presence, prox.city finds and activates the right human.
The human answers by WhatsApp, text, or video. prox.city structures the response, gets approval, and returns the briefing so the agent can continue.
judgment, context, validation, trustWhatsApp, text, or video responseapproved briefing before continuationagent workflow resumes with context
$ route human_gaplive task
agent flow
customer success agent handling enterprise escalation
attempted step
draft response + retention offer
human gap
account owner judgment + negotiation context
activation
WhatsApp first, video if ambiguity remains
return
approved briefing lets the agent continue
agent
paused at gap
human
answer requested
return
briefing pending
activation queue
WhatsAppsent scope + budget
Videooptional intake if risk escalates
Consentrequired before agent execution
right human
prox.city finds who should answer
The system routes the gap to a qualified person based on skill, trust, relationship, availability, constraints, and fit.
live response
The answer comes through a real channel
The human can respond by WhatsApp, text, or video, so the workflow gets usable context instead of a stale profile lookup.
approved return
The agent receives a briefing it can use
prox.city structures the response, preserves approval, and returns the result so the agent can continue the flow.
§01 - Orchestration route
Agent tries to execute. A human-only step appears. prox.city handles it.
The pattern is intentionally simple: detect the human gap, activate the right person, structure the response, get approval, and return the result to the agent flow.
01agent
Agent tries to execute
The workflow is already in motion. The agent is planning, coding, supporting, researching, reviewing, or operating.
02gap
A human gap appears
The next step requires judgment, domain context, approval, trust, negotiation, validation, or presence.
03route
prox.city activates the right human
The system selects who should answer and prepares the smallest useful context for that person.
04channel
Human responds by text or video
The response can happen through WhatsApp, text, or a video call depending on urgency and ambiguity.
05structure
System structures the response
prox.city turns the human answer into fields, constraints, decisions, risks, and next actions.
06approve
Human or company approves
The consequential part stays explicit: who approved, what was approved, and what the agent is allowed to do.
07continue
Agent continues
The agent receives the approved briefing and continues the workflow with human context attached.
§02 - Operator console
The channel captures the human answer. The product returns the agent-ready briefing.
WhatsApp, text, and video are just the activation surfaces. prox.city owns the routing, context, response structure, approval, and continuation handoff.
activation event log
14:02agentattempted to continue support workflow
14:03systemdetected human-only negotiation step
14:04prox.cityselected account owner and sent WhatsApp context
14:09humananswered with constraint, offer limit, and risk note
14:10systemstructured response into approved briefing
14:12companyapproved handoff and continuation boundary
14:13agentresumed workflow with human context attached
human gap analysis
route
supplier risk review
The agent can draft, but a trusted human must validate local procurement risk.
budget
R$ 4,500 cap
Human work is bounded before outreach. Payment terms become part of the handoff.
channel
WhatsApp first
Fast confirmation happens async; video is reserved for ambiguous or high-risk context.
gate
approval required
No agent execution starts until scope, consent, and review criteria are approved.
§03 - Approved handoff
The output is not a message. It is an approved briefing.
The agent gets back a structured object: what the human said, what was approved, what risk remains, and what the agent is allowed to do next.
agent
Blocked workflow
attempted action
reason for gap
needed decision
continuation target
human
Human response
judgment
context
constraints
approval boundary
handoff
Approved briefing
structured answer
owner and timestamp
risk note
agent next action
§04 - Application scenarios
Six places where agents need a human handoff.
The category is not networking. It is bounded human activation inside an agentic flow whenever the next step needs responsibility, trust, context, or presence.
Enterprise operations
An agent needs approval or internal context before continuing an operational workflow.
Software / product teams
An agent codes or plans, but needs technical review, design, QA, copy, domain input, or product decision.
Customer support / success
An agent handles a customer, but needs a human specialist with a complete briefing.
Field operations
An agent needs real-world action: inspection, photo, delivery, installation, or physical verification.
Legal / compliance / finance
An agent prepares analysis, but a human validates, approves, or assumes responsibility.
Research / due diligence
An agent gathers data, but needs an expert interview, source validation, or human judgment.
§05 - Workflow pilot
Give us one agent workflow that keeps stopping for a human.
We will map the human-only step, define the activation channel, structure the approval gate, and show what the continuation handoff should look like.
1 task1 human gap1 approved handoff