LangGraph or a durable workflow engine?

Both, usually. LangGraph holds the reasoning shape. A durable engine holds the multi-system state, retries, and compensation. Using the graph as the system of record for a week-long process is how weekends get ruined.

Is this a replacement for Agentforce?

No. It is a peer for the steps Agentforce should refuse. If the work and the permissions live in the org, start there.

How do you keep a cyclic agent from spending forever?

Hop limits in the graph, budgets on the gateway, and an alert when a run exceeds the expected path. All three, not one.

AI stack

A graph that can loop, and a gate that can stop it.

The first week is in the org, reading what already exists. LangGraph fits the work that is not a straight line: retry, branch, ask a specialist, and pause for a human before the expensive hop. The pause is the feature. The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.

Send the brief

What we put in

Supervisor, workers, and an interrupt.

01
Model the graph
Nodes, edges, and state written down. A graph is reviewable in a way a chain of prompts is not.
02
Supervise
One node decides which worker runs next, so specialists stay small and testable.
03
Loop with a budget
Retries and branches are explicit, with a hop limit. An agent that can loop forever will.
04
Interrupt for a person
A deterministic pause before the write, resolved on the surface the approver already reads.
05
Bound it
For multi-system, long-running work, the graph sits inside a durable workflow that can retry and compensate.

A LangGraph shape we build often. A supervisor node routes to specialist workers, the graph loops while checks fail, and a human interrupt blocks the high-impact write until someone approves it on a surface they already use.

Mermaid source

flowchart TD
  A["Entry: event, message, or schedule"] --> S["Supervisor node: route the step"]
  S --> W1["Worker: retrieve and ground"]
  S --> W2["Worker: compute or transform"]
  W1 --> V["Checks: schema, policy, confidence"]
  W2 --> V
  V -->|"retry"| S
  V -->|"needs a person"| H["Interrupt: approval in Slack or the org"]
  H --> D["Durable write"]
  V -->|"clear"| D

When it earns the seat

Cycles, specialists, and approval gates.

Governed

Cycles do not get their own security model.

Questions

What we actually say.

LangGraph or a durable workflow engine?: Both, usually. LangGraph holds the reasoning shape. A durable engine holds the multi-system state, retries, and compensation. Using the graph as the system of record for a week-long process is how weekends get ruined.
Is this a replacement for Agentforce?: No. It is a peer for the steps Agentforce should refuse. If the work and the permissions live in the org, start there.
How do you keep a cyclic agent from spending forever?: Hop limits in the graph, budgets on the gateway, and an alert when a run exceeds the expected path. All three, not one.

What to read next.

The brief

Name the step that keeps looping.

We will tell you whether it needs a graph, a queue, or a decision nobody has made yet.

A graph that can loop, and a gate that can stop it.

flowchart TD A["Entry: event, message, or schedule"] --> S["Supervisor node: route the step"] S --> W1["Worker: retrieve and ground"] S --> W2["Worker: compute or transform"] W1 --> V["Checks: schema, policy, confidence"] W2 --> V V -->|"retry"| S V -->|"needs a person"| H["Interrupt: approval in Slack or the org"] H --> D["Durable write"] V -->|"clear"| D