State & Persistence
Where does agent state live?
What happens when a process restarts or another worker handles the next request?
In development · Early access
Your AI Agent Works Locally.
Now make it survive production.
AI coding tools can help you build an agent surprisingly fast.
The hard part starts when that agent has to run reliably outside your laptop.
Join Early Access
Free to join · Help shape the first version
- $ python -m agent --env local
- ✓ 12/12 runs completed
- $ deploy --env production
- ✗ llm.invoke: timeout after 30s
- ↻ retry 1/3 · send_invoice(order=4821)
- ✗ send_invoice executed twice
- ⚠ client disconnected · run 7f3a still active
- ✗ checkpoint missing on worker-2
- ? why did the agent call refund_order
The hard part starts outside your laptop
- What happens when an LLM times out?
- When a tool fails halfway through execution?
- When a retry executes the same action twice?
- When the client disconnects?
- When you need to understand why the agent made a particular decision?
That's the gap we're exploring at Salada de Dados.
From Working Prototype to Production System
We're building a practical resource for developers taking AI agents beyond the prototype stage.
Not another introduction to agents. Not another collection of copy-and-paste code.
Instead, we'll focus on the engineering problems that appear when an agent meets production.
Failures & Retries
LLMs fail. APIs time out. Tools return unexpected results.
Which operations can safely be retried — and which cannot?
Idempotency
What happens when the same tool executes twice?
A duplicate weather request may not matter.
A duplicate payment, email, order, or database mutation certainly does.
Observability
Your agent produced the wrong result. Now answer the harder question: Why?
What model calls happened? Which tools ran? What state changed? Where did the execution fail?
Streaming & Disconnects
Agent requests can take time.
What happens when the browser disconnects halfway through execution? Should the agent stop? Should it continue? What happens to its state?
Built for AI-Assisted Development
We assume you're already using AI to help write code.
The goal isn't to compete with your coding assistant.
The goal is to help you ask better engineering questions before generated code reaches production.
For each production problem, we're exploring a practical workflow:
- 01Failure Scenario
- 02Risk
- 03Architecture Decision
- 04Pattern
- 05Implementation
- 06Test
Along with:
- Reference implementations
- Production checklists
- Prompts to review your own code with AI
What We're Building
The first version is planned around five production problems.
The scope may change.
That's why we're opening Early Access before building everything.
State & Persistence
Designing agent state that survives beyond a single request.
Failures & Retries
Handling unreliable models, tools, and external APIs.
Idempotency
Preventing retries and duplicate execution from causing unintended actions.
Observability
Understanding what the agent actually did during execution.
Streaming & Disconnects
Handling long-running requests and interrupted clients safely.
Help Shape the First Version
If you're building AI agents, join the Early Access list. You'll get updates as we build and test the production patterns.
More importantly, tell us where you're getting stuck.