HOME/TECHNIQUE/Agentic Orchestration/Human-in-the-loop checkpoints

TECHNIQUE

Human-in-the-loop checkpoints

Agentic Orchestration

12APPLICATIONS
12OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 5 OPERATORS

Observed practice: operators keep human checkpoints around agent outputs that become production changes, PR changes, governance approvals, or published incident artifacts.

Observed Practices

Require a platform-team review before a new AI use case is onboarded: Grab uses a mini-RFC and checklist for every new use case.

1 of 5 observed operators with cited human-checkpoint evidence.
Grab

Keep human approval before production deployment of generated security changes.

1 of 5 observed operators with cited human-checkpoint evidence.
Amazon

Route AI-generated code fixes or optimization suggestions through developer review / PR workflows rather than applying them silently.

2 of 5 observed operators with cited human-checkpoint evidence.
UberDoorDash

Validate AI-drafted incident/report content with a human reviewer before publishing internally.

1 of 5 observed operators with cited human-checkpoint evidence.
Agoda

Escalate to human review when the LLM lacks full context, such as prior alerts or historical signals.

1 of 5 observed operators with cited human-checkpoint evidence.
Agoda

Use automated validation or falsification before the human checkpoint to reduce low-quality outputs reaching reviewers.

3 of 5 observed operators with cited human-checkpoint evidence.
UberDoorDashAmazon

Where Operators Converge

Across the cited deployments, human checkpoints are retained around durable or high-impact outcomes: use-case onboarding, production security changes, PR/code changes, or published incident documentation.

Where Operators Diverge

Operators place the human checkpoint at different stages of the workflow.

APPROACH 01

Upfront governance gate before a use case is onboarded.

Grab

APPROACH 02

Approval gate before deploying generated security changes to production.

Amazon

APPROACH 03

PR or developer-review checkpoint for AI-generated code fixes and optimization findings.

UberDoorDash

APPROACH 04

Post-generation reviewer validation before publishing an AI-drafted incident report.

Agoda

Operators differ on whether human involvement is mandatory for every item or conditional on uncertainty / interaction.

APPROACH 01

Mandatory review or approval is built into the workflow.

GrabAmazonUberAgoda

APPROACH 02

Human review is used as a fallback when context is insufficient.

Agoda

APPROACH 03

Humans can invoke the agent from the PR thread to request a change.

DoorDash

Watch Items

False positives, hallucinations, and weak findings are recurring reasons for validation before or alongside human checkpoints.

Insufficient context is explicitly treated as a reason to avoid autonomous resolution and route to people or further verification.

For production/security-sensitive changes, operators do not present the agent as the final authority; approval remains a human checkpoint.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
Approval queue with resumable statepatternestablished
LangGraph interruptslibraryestablished
Temporal signalsserviceestablished
03

Observed in Production

12 APPS