Workflow to agent mapping
We shadow the work, then turn it into a spec the agent can follow: inputs, decisions, tools, limits and hand-off points.
Discipline01 · Agent Engineering
In shortUpdated 30 Oct 2026 · Labs 24 FDE Squad
AI agent development at Labs 24, in short
Agent engineering is the discipline of designing, building and running AI agents that take multi-step actions inside real business systems. Labs 24 does AI agent development with forward deployed engineers who embed in your team, build on your live data and ship into your infrastructure.
- What we build
- Triage, back-office, document, channel, analyst and dispatch agents
- How we ship
- One slice live on real work, then the next
- Safety
- Autonomy set per action (L1 to L4), approvals in code, evals, full tracing
- Where it runs
- Your cloud or on-prem, including self-hosted open-weight models
- Stack
- LangGraph, Temporal, n8n, MCP, Claude, GPT, Gemini, Llama, Qdrant, Langfuse
- Pricing
- Set after an initial embed; operations agents priced against savings
01 / What it means
An agent is a job description, then code.
Most agent projects stall because the agent was designed from a whiteboard. Ours start from watching the person who does the work today: what they read, which systems they open, where they make a judgement call, and when they escalate.
First: we watch the work before we write the agentAgent built from a brief
- Broad goal: "handle customer emails"
- Tested on clean sample data
- Same autonomy for every action
- Failures found by customers
Agent built forward deployed
- One named job with a clear finish line
- Built and tested on your live data
- Autonomy set per action, by risk
- Every run traced and scored by evals
- A named owner on your side from day one
02 / Capabilities
AI agent development capabilities.
Eight parts of agent engineering we own end to end. Most deployments use five or six of them.
Tool and API design
Typed, narrow tools the agent can call safely. We wrap your ERP, CRM and internal APIs, and expose them through MCP servers where useful.
Orchestration
Stateful agent graphs, retries, timeouts and long-running jobs that survive restarts. Single agents or multi-agent teams, picked per job.
Human in the loop
Approval steps, escalation rules and review queues in the tools your team already uses, so people stay in charge of the calls that matter.
Memory and context
Short-term state per task, long-term memory per customer or case, and retrieval over your documents, scoped so agents only see what they should.
Guardrails and policy
Spend limits, allow-lists, PII redaction, output validation and hard stops, enforced in code instead of in the prompt.
Channel agents
Agents that work where your customers and staff already are: WhatsApp, SMS, email, voice and internal consoles.
Evals and observability
Every run traced, scored and replayable. Regression evals run before each release, so a prompt change never quietly breaks last month's work.
03 / An agent run
Watch one claim go through.
An example run from a claims-triage agent. It reads the email, checks the order, drafts the credit note and stops for a person because the amount is over its limit.
- 00.0sRead inbound email
ops@· "Box crushed, order 88213"Done - 00.8sClassify request Damaged shipment · confidence 0.94Done
- 01.4sCall
tms.lookup_orderDelivered 2 days ago · POD photo on fileDone - 02.1sDraft credit note $480 against invoice INV-30917Done
- 02.2sAsk for approval Over the $250 auto-approve limit · approved by Ops deskApproved
- 41.6sPost credit note to ERP Ledger updated · audit entry writtenDone
- 42.3sReply to customer Apology and credit confirmation sentDone
7 / 7Steps
$0.011Model cost
L2Autonomy
What you get on every run
A full trace
Every input, tool call and decision, replayable step by step.
A person where it matters
Approval rules sit in code, so the agent cannot skip them.
Cost and speed per run
Model spend and latency tracked against the baseline from the start.
Safe changes
New prompts and models are tested against past runs before release.
04 / Autonomy
How much an agent may do on its own.
We set autonomy per action, not per agent. The same agent can refund $50 on its own and stop for sign-off on $5,000.
L1
Suggest
Drafts the answer or the action. A person reviews and sends it.
L2
Act with approval
Prepares everything and waits for one click before it changes a system.
L3
Act and report
Acts inside set limits and logs every step for review.
L4
Run the queue
Owns a whole queue and only escalates exceptions to people.
| Agent pattern | Takes over | Typical tools | Usual start |
|---|---|---|---|
| Triage agent | Sorting, tagging and routing inbound emails, tickets and claims | Inbox, helpdesk, CRM | L2 → L3 |
| Back-office agent | Reconciliations, order exceptions, credit notes, vendor follow-ups | ERP, billing, spreadsheets | L2 |
| Document agent | Reading contracts, invoices and forms, then extracting and checking them | DMS, OCR, vector store | L1 → L2 |
| Channel agent | Customer and field conversations on WhatsApp, SMS, email and voice | Messaging APIs, order systems | L3 |
| Analyst agent | Pulling data across systems and writing the weekly report or brief | Warehouse, BI, docs | L1 |
| Dispatch agent | Assigning jobs, chasing confirmations, collecting proof of delivery | TMS, WhatsApp, maps | L3 → L4 |
05 / How it ships
From shadowing to production, one slice at a time.
The same build order as every Labs 24 deployment, applied to an agent. A demo every Friday on your real cases, until one slice runs in production.
Engineers on your floor from day oneStep 01 · Embed
Shadow the work
We sit with the team, log every step of the workflow and baseline how long it takes today.
- Job spec for the agent
- Tool list and access
- Autonomy level per action
Step 02 · Build
First working run
Tools wired to your systems and the agent running end to end on real cases in a sandbox.
- Typed tools and guardrails
- Tracing switched on
- Friday demo
Step 03 · Prove
Run beside your people
The agent works the same queue as your team in shadow mode. We compare its output to theirs.
- Eval suite from real cases
- Approval flow in your tools
- Go-live checklist
Step 04 · Live
One slice in production
The agent takes live work on one slice of the queue, with a named owner on each side.
- Live dashboard and alerts
- Runbook and handover
- Plan for the next slice
06 / Stack
AI agent frameworks and stack.
Model-agnostic by design. We choose the model per step on cost, speed and accuracy, and can swap it later without a rebuild.
Orchestration
LangGraphTemporaln8nMCP serversAgent SDKs
Models and memory
ClaudeGPTGeminiLlamaQwenvLLMQdrantNeon
Run and observe
LangfuseEval suitesAlertingYour cloudOn-prem
Need data to stay inside your network? We run open-weight models on your own hardware. See Production & Cloud Engineering.
07 / In production
AI agents in production.
Under NDA, so clients stay unnamed. Each one runs on the client's own infrastructure with full IP transfer.
Regional dispatch agent
Assigns jobs, chases drivers on WhatsApp and collects proof of delivery photos back into the TMS.
Dispatch · L3
Patient follow-up agent
Runs post-visit follow-ups for a clinic chain and routes anything clinical straight to staff.
Channel · L3
Law practice document agent
Reads and checks case documents on self-hosted models, so client files never leave the firm.
Document · L1
08 / The squad around it
No agent ships with one skill set.
A production agent needs the model tuned, the data clean, the systems connected and the risk signed off. These disciplines sit in the same squad.
02
LLM Engineering
Model choice, prompts, structured outputs and cost per run.
03Data & Retrieval
The documents and records the agent reasons over.
04Evals & Reliability
Proof the agent still works after every change.
05Production & Cloud
Deployment, scaling and self-hosted models.
06Integration
The ERP, CRM and TMS connections behind every tool.
11Security & Compliance
Access control, PII handling and audit trails.
09 / FAQ
AI agent development FAQ.
What is agent engineering?
Agent engineering is the discipline of designing, building and running AI agents that take multi-step actions inside real business systems. An agent reads inputs, calls tools and APIs, makes decisions within set limits, and hands off to people when needed. Labs 24 builds these agents with forward deployed engineers who work inside the client's operation.
What does an AI agent development company do?
An AI agent development company maps a workflow, designs the tools the agent can use, connects it to systems like the ERP, CRM or helpdesk, sets guardrails and approval rules, tests it against real cases and runs it in production. Labs 24 does this with forward deployed engineers embedded in the client's team.
How is an AI agent different from an automation or a chatbot?
An automation follows a fixed path. A chatbot answers questions. An AI agent decides which steps to take, uses your tools to take them, and finishes a piece of work, with human approval wherever the risk calls for one.
What is the difference between agentic AI and an AI agent?
Agentic AI is the broader approach of letting AI plan and act toward a goal. An AI agent is one concrete system built that way, with a defined job, a set of tools and limits on what it may do on its own.
How do you stop an AI agent from doing the wrong thing?
Each action gets an autonomy level from L1 (suggest) to L4 (run the queue). Tools have hard limits in code, high-impact actions need human approval, an eval suite built from real cases runs before every release, and every run is traced.
Which AI agent frameworks do you use?
Labs 24 picks per job: LangGraph for stateful agent graphs, Temporal for long-running durable workflows, n8n where operations teams need to see and edit the flow, MCP servers for tool access, and model-provider SDKs where they fit. Designs stay model-agnostic across Claude, GPT, Gemini and open-weight models.
Can AI agents run on our own infrastructure?
Yes. Labs 24 deploys into the client's cloud or on-prem environment and can run open-weight models such as Llama or Qwen through vLLM or Ollama when data cannot leave the network. Clients get full IP transfer.
What are the steps in building an AI agent?
Labs 24 shadows the work and writes the agent's job spec, wires its tools to your systems and runs it on real cases in a sandbox, runs it in shadow mode beside your team, then puts one slice of the queue live with a named owner. There is a demo every Friday throughout.
How much does AI agent development cost?
Labs 24 prices after an initial embed, once the current work is baselined. Operations agents are priced against the time and cost they save, so the price is tied to a measured outcome.