Regulated support · agent with human escalation
Agent handles 78% of tickets end to end — opens, resolves, closes — and escalates with full context on the remaining 22%. Every decision has an audit trail.
Systems that act on behalf of your business.
Agents are not chatbots. They are systems that connect data, tools and processes — with supervision, metrics and predictable cost.
What we do
Agents that close the loop, from intent to action.
Every agent has explicit tools, declared scope, human supervision and quality metrics. All auditable.
Customer-facing agents
Autonomous resolution with human escalation. Voice and text. Integrated with CRM, ERP and internal knowledge.
Qualification agents
Lead triage and prioritization, with enrichment and routing by intent and fit.
Operations agents
Cognitive backoffice: document validation, accounting classification, contract review.
Analytics agents
On-demand analysis over data warehouses. Natural-language answers with audit-ready SQL.
Human supervision
Approval loops, doubt instrumentation, corrections feeding evaluation.
Security and isolation
Sandbox, tool scoping, redaction, data residency, signed logs.
Thesis
An agent is not a chatbot. The gap is worth 10×.
Chatbots answer — agents act. Chatbots stay trapped in the conversation; agents close the loop in real systems. When you measure time saved per interaction, chatbots return 1-2 minutes; agents return 15-30. One pivots the whole operation, the other leaves a report.
By the numbers
What shifts once the agent reaches production.
−0%average handle time
0%auto-resolution with audited quality
<0msmedian latency per turn
0%of actions logged and signed
Anatomy
From intent to auditable action.
Every production agent has declared tools, data scope, human supervision and quality metrics — all auditable at runtime.
- A1
Intent
Intent classifier + confidence routing + human fallback for ambiguity.
Output Top-3 intents with audited accuracy.
- A2
Tools
Finite declarative set of executable actions (create order, open ticket, update record, escalate).
Output Versioned tool registry with schemas.
- A3
Memory
Short context (session) + long memory indexed by user/account. Automatic PII redaction.
Output Vector store with per-user ACL.
- A4
Supervision
Human-in-the-loop on sensitive actions. Doubt/correction instrumentation feeds evaluation.
Output Supervision panel + review queue.
- A5
Observability
Signed logs, cost per turn, latency, auto-resolution rate, semantic drift. SLO alerts.
Output Dashboards with per-agent SLO.
Under the hood
The stack behind each agent.
Vendor-neutral by principle — we pick model, framework and infra by criteria, not prior preference. The composition below is the most common across production deployments.
LLMs
- Claude (Anthropic)Default for reasoning + structured tool use.
- GPT (OpenAI)Where Brazil region is available.
- Gemini (Google)Multi-modal and long context.
- Self-hosted open-sourceLlama / Mistral for regulated data.
Frameworks
- LangGraph / LangChainOrchestration + state machines.
- Pydantic AITyped tool use.
- Vercel AI SDKStreaming UX on the frontend.
Memory + RAG
- pgvector / PineconeVector store per use case.
- BM25 + rerankingHybrid for quality.
- Cohere rerankVerifiable citations.
Observability
- Langfuse / PhoenixEnd-to-end tracing.
- OpenTelemetryStandard for the rest of the stack.
- Custom eval suiteGold-set + drift.
In production
Agents that close the loop.
B2B qualification · pre-sales agent
Lead triage and enrichment with intent and fit-based routing. SDRs only get leads with score >70 and ready context for the first call.
Cognitive backoffice · operations agent
Document validation, accounting classification and contract review with human-in-the-loop. Reduced rework and standardized output across branches.
Frequently asked
About agents in production.
What's the difference between agent and chatbot?
Chatbots answer. Agents do. A chatbot delivers information in a conversation; an agent executes actions in real systems (creates an order, opens a ticket, updates a record) and closes the loop. Business value is typically 10× different.
How do you ensure the agent doesn't act outside its scope?
Each agent has explicit tools declared (a finite list of actions it can call), data scope (what it can read and write) and human supervision on sensitive actions. Everything logged with signed payload.
Does it work with our legacy systems?
Yes. We connect via existing APIs, RPA when there's no API, and in extreme cases via screen scraping with guardrails. The agent stays decoupled from the underlying system — replacing the backend doesn't break the agent.
How much does an agent cost to run?
Cost is modeled per interaction, not per user. Our deployments average between R$ 0.02 and R$ 0.15 per full turn, depending on complexity. Compared to human handle cost, it pays itself quickly.