# Basalt

Melhore agentes de IA analisando traces reais, detectando falhas e gerando pull requests validados.

- Findings24: <https://findings24.com/pt/tools/basalt>
- Site: <https://getbasalt.ai>
- Preços: Freemium
- Categorias: Assistente de Código IA, Observabilidade de IA, Assistentes de testes com IA

## O que é Basalt?

Basalt is a product for improving AI agents from real user interactions. It connects to observability or trace sources, reads the previous day’s conversations, and identifies recurring failure patterns that point to where an agent’s behavior needs adjustment.

When it finds a problem, Basalt generates a targeted pull request that can change a prompt, tool definition, or system instruction. Before the PR is opened, the proposed fix is validated against a golden dataset so existing behavior is checked for regressions.

## O que Basalt pode fazer?

### Trace ingestion from observability tools

Connects to Langfuse, Braintrust, Datadog, or a custom trace source and ingests conversation data automatically on a nightly cadence.

### Failure-pattern analysis

Uses Claude to cluster conversations, detect failure patterns, and identify which interactions point to quality gaps.

### Automated fix generation

Generates a focused PR that can update a prompt, tool definition, or system instruction based on the detected issue.

### Golden-dataset validation

Runs each generated fix against a golden dataset before the PR is opened, blocking changes that introduce regressions.

### Behavior-driven workflow

Positions the product as an automated review loop that does not rely on LLM-as-a-judge or manual annotation.

### Daily improvement cycle

Sends one high-quality PR every morning from the previous day’s user behavior, keeping the workflow tied to recent interactions.

## Casos de uso

### Review production conversations

Teams that already record traces in observability tools can connect those sources and let Basalt scan conversations overnight for recurring errors.

### Patch repeated failure patterns

When the same misunderstanding or tool-use mistake keeps appearing, Basalt can generate a focused PR that targets the affected behavior pattern.

### Protect existing behavior

Teams that maintain golden datasets for expected behavior can use Basalt’s validation step to check that a proposed fix does not regress known good cases.

### Draft agent improvements

Organizations that want prompt or tool updates without hand-writing every change can use Basalt to propose modifications based on real user behavior.

### Triage changes in a daily review cycle

Teams looking for a lightweight improvement loop can review one daily PR rather than manually sorting through every interaction.

## Perguntas frequentes

### How does Basalt turn user interactions into an update?

Basalt ingests conversations from connected observability sources, analyzes them overnight, and turns repeated failure patterns into a targeted PR. The homepage says it can generate prompt updates, tool definition changes, or system instruction changes, then validate the fix against a golden dataset before opening the PR.

### What trace sources does Basalt support?

The homepage explicitly names Langfuse, Braintrust, Datadog, and a custom trace source as connection options. It presents Basalt as working on traces from those sources rather than requiring manual annotation.

### What parts of an agent can Basalt improve?

Basalt is positioned around improving agent behavior from real user conversations. The page highlights system prompts, tool definitions, and few-shot examples as the parts it can change when it finds a failure pattern.

### How does Basalt check that a fix does not break existing behavior?

The homepage says Basalt runs fixes against a golden dataset before a PR is opened and blocks the PR if regressions appear. It also states there is no LLM-as-a-judge and no manual annotation in the main workflow.

### How is Basalt priced?

The pricing section says you only pay when you merge a PR, with the first 3 PRs free and no subscription or commitment. It also notes pricing applies per merged PR per repository and that volume discounts are available for teams merging 20+ PRs per month.

## Informações rápidas

- **Category**: AI developer tool
- **Primary use**: Improving AI agents from user interactions
- **Workflow**: Nightly trace analysis and one PR per morning
- **Supported sources**: Langfuse, Braintrust, Datadog, and custom trace sources
- **Pricing shape**: Pay per merged PR; first 3 PRs free
- **Source domain**: getbasalt.ai