LangGraph logo

[ AI tools · Frameworks and platforms ]

LangGraph Agents That Wait, Resume and Explain

LangGraph is our default framework for agents that run for a long time, pause for human approval or need an audit trail. Every step is a node we can test, and every state is saved.

Design Your AgentSee a Resume
Read invoiceMatch to POCheck rulesHuman approvalPost to ERPNotify supplier

Running

checkpoint 00 saved

Example invoice workflow. The run can wait for days at the approval node and resume where it stopped.

Every Step Saved, Every Step Replayable

LangGraph builds agents as graphs whose state is saved after each step, in Python or TypeScript, with any model. Drag the slider to rewind an example run.

thread invoice-2207 · checkpoint 5 of 5Completed
state = {
  "step": "post_to_erp",
  "note": "Posted to the ERP, receipt streamed to the finance channel"
}
What this shows

Streaming

Tokens, steps and custom events streamed to your interface as the agent works.

Example run, simplified. LangGraph also works in TypeScript with any model provider.

[ Durability ]

When Something Breaks, the Run Picks Up Where It Stopped

The manager approves in the morning; the run continues and posts the invoice.

[ How we build ]

Five Parts of Every LangGraph Agent We Ship

Each part is small and testable on its own, which is what makes a long-running agent safe to change later.

  1. 1

    State

    A typed object every node reads and updates: the invoice, the match, the decision, the audit notes.

  2. 2

    Nodes

    Plain functions. Some call a model, others call your ERP or apply rules in code.

  3. 3

    Edges

    What happens next, including conditional edges chosen by the model or by a rule.

  4. 4

    Interrupts

    Points where the graph stops and waits for a person, with the state shown to them.

  5. 5

    Checkpointer

    Where state is saved after each node, so every run is resumable and reviewable.

Manager reviewing a document on a clipboard at his desk
INV-2207 needs your approval01 / 04INV-2207 approved: within supplier historyAmount corrected to 4,108 EUR, run resumedSent to finance lead with full contextPaused 3 days, resumed at step 6
[ Human in the loop ]

The Approver Sees What the Agent Saw

When a run pauses, we show the person the state behind the decision, not just a yes or no button.

  1. 01

    Approve or reject

    With a reason that is stored in the run's audit trail, so the decision can be explained later.

  2. 02

    Edit before continuing

    Correct an extracted field and let the run carry on with the fix.

  3. 03

    Escalate

    Send it to someone with higher limits without losing the context the agent gathered.

  4. 04

    Wait as long as needed

    Runs pause for hours or weeks without holding a server, then resume from the same step.

[ In production ]

What Runs Around the Graph

  • Hosting

    LangSmith Deployment, or your own containers on AWS, Azure or Google Cloud.

  • Tracing

    LangSmith or Langfuse traces for every node, with cost and latency.

  • Evaluation

    Saved test runs scored on every change to prompts, models or graph.

  • Tools

    MCP servers and typed tools with scoped credentials for your systems.

Choose a Starting Point

Packages from $1,500; custom projects from $8,000.

  • One LangGraph agent with approvals, checkpoints and tracing, connected to your tools.

    • Graph designed with your team
    • Postgres checkpoints
    • Approval step for risky actions
    • Traces in LangSmith or Langfuse
    Details

  • A senior engineer who builds and maintains LangGraph agents in your team.

    • Starts within 48 hours
    • Works in your repositories
    • Reviews and tests on every change
    Details

  • Several agents on shared state, approvals, evaluations and deployment.

    • Multiple graphs on shared state
    • Evaluation suite in CI
    • Deployment in your cloud
    • Runbook and handover
    Scope it

LangGraph Questions

  1. 1

    When should we use LangGraph?

    When an agent runs for more than one step and you need control: approvals, retries, branching, long waits or an audit trail. For a single model call with a tool or two, a provider's SDK is simpler.

  2. 2

    LangGraph or CrewAI?

    CrewAI is quick for role-based agent teams in Python. LangGraph gives finer control over state, checkpoints and human approvals, which matters for regulated or long-running work. We use both and pick per project.

  3. 3

    Do we need LangSmith?

    No. LangGraph is open source and runs anywhere. LangSmith adds hosting, tracing and evaluation; Langfuse or your own observability stack can do the tracing part instead.

  4. 4

    Can LangGraph use any model?

    Yes. Nodes can call OpenAI, Anthropic, Google, open models or several of them in one graph.

  5. 5

    How do approvals work in practice?

    The graph hits an interrupt, saves its state and stops. Your app shows the approver the relevant state; their decision resumes the run from that checkpoint, minutes or weeks later.

  6. 6

    Is it hard to maintain?

    Less than a hidden agent loop, because each node is a small function with its own tests. We hand over the graph diagram, tests and runbook so your team can change it.

Ask Us Directly
[ Contact ]

Let's build your next product together

Book a free strategy call and leave with a clear plan and estimate. No commitment.

  1. 01Pick a time that suits you
  2. 0230 minutes on scope, stack, timeline and budget
  3. 03Fixed-price proposal, NDA on request
Prefer email? contact@sajaltech.com