Single agent
One concrete workflow end to end: classifying and routing requests, preparing quotes, reconciling documents, following up on a process.
4–6 weeks
AI agents
Agents that do the work, not just talk about it
A chatbot answers questions. An agent queries your systems, decides, executes the action, and knows when to ask a human for approval.
Technical capabilities
LangGraphMulti-agent orchestrationTool useMCPHuman-in-the-loopTraceability
4–6weeks from workflow to first agent in production
2M+users served by systems we have operated
100%of sensitive actions behind configurable human approval
Key takeaways
- An AI agent differs from a chatbot in that it executes: it takes a goal, decides the steps, uses tools across company systems and corrects course.
- Quarl builds agents on LangGraph, with durable state, human approval on steps that carry consequences, step and cost limits, and per-run traceability.
- Formats range from a single agent to coordinated multi-agent systems, internal copilots, and Model Context Protocol servers that expose systems as standardized tools.
- The first agent reaches production in four to six weeks, preceded by a one-to-two-week exploration that avoids building on assumptions.
Operating track record
FEMSA loyalty platform
Three years operating the system in production — not one delivery and an exit.
Cross-entity contamination
The most frequent failure in RAG architectures, solved with structured metadata at ingestion.
Fixed scope and date
The proposal arrives within 48 hours with fixed scope, price and date. If scope changes, it is quoted separately and approved first.
What an agent is
The difference between answering and executing
A conversational assistant retrieves information and writes it up. An agent does something more: it receives a goal, decides which steps to take, uses tools — your CRM, your ERP, your database, an external API — evaluates the result and corrects course when needed.
That autonomy is exactly what makes it useful and what makes it dangerous. A badly built agent executes wrong actions quickly and silently. That is why everything we build carries explicit guardrails: scoped permissions, human approval checkpoints on sensitive actions, and an audit trail of every step.
Types of agent
What we build
Multi-agent systems
Several specialized agents that coordinate and hand work to each other, with an orchestrator deciding who does what and when to escalate.
8–12 weeks
Internal copilots
An agent inside the tool your team already uses, with operational context and role-based permissions.
6–10 weeks
Custom MCP servers
We expose your systems as standardized tools through Model Context Protocol, so any agent can use them without brittle one-off integrations.
2–4 weeks
What separates an agent from a demo
Five things almost nobody implements
- Durable state. The agent survives restarts, resumes where it stopped and does not repeat actions already executed. Without this, any failure means redoing the work from scratch.
- Human in the loop. Actions that move money, write to systems of record or talk to a customer go through configurable approval.
- Tool boundaries. Each agent reaches only what its task requires, with role-based permissions and tenant isolation.
- Retries and replay. When a tool fails the agent retries sensibly; when something goes wrong the run can be replayed step by step to understand what happened.
- Cost and step limits. An agent without a ceiling loops and burns budget. Ours all carry step, time and spend caps.
Process
From workflow to agent in production
- Map the real workflowWith the person who runs it today, not with the org chart. We routinely find three undocumented exceptions — and those are exactly what break badly built agents.
- Design the graph and the guardrailsWhich steps are automatic, which need approval, which tools get exposed and with what permissions. Agreed before any code is written.
- Build with evaluation runningA set of real cases with expected outcomes, running from day one. Without it there is no way to know whether the agent is improving or degrading.
- Scoped pilotReleased against a real subset of the work, supervised and measured. It expands when the numbers justify it, not when the demo impresses.
- Production and observabilityStep-level traces, failure and cost alerts, and a dashboard of what the agent did. Delivered documented.
Pricing
Ranges and timelines
| Scope | What it includes | |
|---|---|---|
| Exploration | Workflow review, feasibility, graph design and estimate. Enough to decide with data whether this is worth it. | 1–2 weeks |
| Single agent in production | One complete workflow with tools, guardrails, evaluation and observability. | 4–6 weeks |
| Multi-agent system | Coordinated agents, orchestrator, approvals and full traceability. | 8–12 weeks |
| Support and evolution | Maintenance, prompt and tool tuning, cost and quality monitoring. | Monthly |
Exploration is credited against the project if we move forward.
Stack
What we build agents with
LangGraphLangChainModel Context ProtocolOpenAIAnthropic ClaudeGoogle GeminiAzure OpenAIAmazon BedrockLangSmithPydanticFastAPINestJSPostgreSQLRedisDockern8n
Related services
LangChain and LangGraph
Implementation and observability instrumented with LangSmith.
View serviceRAG systems
Chunking, reranking and hybrid search, evaluated with recall@k and NDCG.
View serviceAI automation
Applied where there is volume and stable rules, not where there is expectation.
View serviceAI consulting
Architecture, evaluation criteria and cost per query before writing code.
View serviceFrequently asked questions
What is the difference between an agent and a chatbot?
A chatbot converses: it takes a question and returns text. An agent executes: it takes a goal, decides which steps to take, uses tools to query or modify your systems, evaluates the result and corrects. A chatbot tells you an order’s status; an agent notices the order is late, notifies the customer, reschedules the shipment and logs all of it.
Which framework do you use and why?
LangGraph for most production cases, because it models the flow as a graph with durable state, allows pausing for human approval, retrying a single step and replaying a full run for debugging. For simple flows n8n with model calls is sometimes enough, and we say so instead of selling complexity.
What if the agent does something wrong?
It is designed so it cannot. Sensitive actions — those that move money, write to systems of record or talk to a customer — go through configurable human approval. The rest is bounded by permissions and by step, time and cost limits. And everything lands in an audit trail, so when something goes wrong you can reconstruct exactly what happened.
Does it connect to our CRM and ERP?
Yes, as long as they expose an API or an accessible database. We have integrated inventory systems, payment platforms, CRMs and in-house databases. When several systems are involved we usually expose them as MCP servers, which standardizes access and prevents every new agent from needing bespoke integrations.
What does it cost to run an agent per month?
It depends on volume and how many steps each run takes. For a mid-market company it usually lands between USD 150 and 900 per month in model and infrastructure consumption. We estimate it with concrete numbers in the proposal, and task-based routing exists precisely to keep that figure low.
Does it replace people?
In practice, almost never. It removes mechanical work — classifying, searching, copying between systems, following up — so the same people handle what requires judgment. Most of our clients automate because their team is underwater. If your goal is headcount reduction, say so upfront: it changes how the project is designed and what has to be documented.
How long until the first agent is in production?
Four to six weeks for a single workflow, counting from when we have access to the information and systems. The prior exploration takes one to two weeks more and exists so we do not build on assumptions.
Can we start small?
That is what we recommend. The exploration reviews one concrete workflow, evaluates whether an agent adds value, and delivers the design with its estimate. If we proceed, it is credited against the project. It is the cheapest way to find out whether this makes sense in your operation before committing a quarter.
Book 15 minutes
Tell us what you are building, or what stopped working. You leave the call with a concrete answer: it can be fixed, it can be built, or it isn't worth it.