Applied AI / agent engineering
Build stateful agents that can pause, recover and ask for judgment.
LangChain tools and retrieval with LangGraph state, checkpoints and human supervision.
We build agent systems around explicit state, typed tools, durable checkpoints and review paths rather than hiding a business workflow inside one model call.
- Delivery
- Stateful
- Scope
- Tool-using
- Control
- Human-supervised
Agent state graph
Durable business outcome- 01Request
- 02Retrieve
- 03Decide
- 04Act
- 05Review
- 06Resume
The operating problem
A model call can answer a question. An operational agent must remember state, choose permitted tools, recover from failure and know when the next step belongs to a person.
System mechanism
The LangChain and LangGraph operating loop
LangChain supplies integrations and retrieval primitives while LangGraph makes workflow state, branching and human intervention explicit.
A request and approved evidence enter explicit workflow state, branch into permitted action or human review, checkpoint progress, recover from failure and finish with an accountable business outcome.
- 01Business inputBusiness request
A bounded process begins with a known result and owner.
- 02LangChain retrievalApproved evidence
Retrieval supplies current, permitted business context.
- 03LangGraph stateExplicit workflow state
The graph records what is known, completed, missing and allowed next.
- 04Conditional edgeBusiness decision branch
Known rules route the run to action, clarification or review.
- 05Typed toolPermitted business action
A typed tool reads, drafts or updates only what this state allows.
- 06Durable checkpointCheckpoint and recovery
Progress persists before costly steps, retries and interruptions.
- 07HITL interruptHuman decision
Uncertainty or consequence pauses with evidence and a proposed next action.
- 08Business valueCompleted business process
The process completes or escalates without losing state, evidence or ownership.
Fatbunny responsibility
Fatbunny defines the state model, tool contracts, retrieval boundary, checkpoint strategy, interrupts, evaluation protocol and interface needed to operate the agent.
Capabilities
- 01
LangChain retrieval and tool integration
- 02
LangGraph state-machine design
- 03
Durable execution and checkpoints
- 04
Human-in-the-loop interrupts
- 05
Agent evaluation and observability
Delivery decisions
- 01
Represent business state explicitly
- 02
Use deterministic code for deterministic rules
- 03
Keep tools narrow and typed
- 04
Checkpoint before expensive or consequential transitions
Expected outputs
- 01
A versioned graph and tool contract
- 02
Working agent interface and human-review path
- 03
Evaluation cases, traces and operational documentation
Honest limits
- !
LangGraph adds value when state and branching are real requirements; simple calls do not need a graph.
- !
Agent memory is not permission to retain arbitrary customer information.
- !
Framework observability does not replace business-outcome measurement.
Why LangChain and LangGraph belong together
LangChain provides useful model, retrieval and tool abstractions. LangGraph provides the durable state and control flow needed when work branches, pauses or resumes.
They solve related parts of one operating problem, so Fatbunny treats them as one delivery ecosystem rather than competing service pages.
State makes behaviour inspectable
When the agent's current state is explicit, a reviewer can see what it knows, which step ran, what failed and what must happen next. This is safer and easier to improve than an opaque loop.
Human intervention is a designed path
A production graph should not merely stop when confidence is low. It should produce the context, proposed action and evidence a person needs to make the next decision and then resume cleanly.
Direct answers
Common questions
What is the difference between LangChain and LangGraph?
LangChain provides model, retrieval and tool integrations; LangGraph provides explicit state, branching, checkpoints and durable workflow control.
Do all agents need LangGraph?
No. It is useful when work has state, branches, retries, pauses or human review. A simple bounded call may need neither framework.
Can LangGraph agents use our internal knowledge?
Yes, through a retrieval layer that respects source quality, permissions, freshness and citation requirements.