Trace ingestion from observability tools
Connects to Langfuse, Braintrust, Datadog, or a custom trace source and ingests conversation data automatically on a nightly cadence.
Basalt
ПодтвердитьBasalt helps teams improve AI agents from real user traces by analyzing failures and generating validated PRs for observability-driven workflows.
Overview
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.
Features
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.
Use cases
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.
Pros and Cons
Pros
- Turns real user behavior into concrete code or prompt changes instead of relying on manual annotation.
- Checks proposed fixes against a golden dataset before opening a PR, which helps catch regressions early.
- Supports multiple kinds of agent updates, including prompts, tool definitions, and few-shot examples.
- Connects to named observability tools and a custom trace source, making it easier to plug into an existing tracing workflow.
- Uses the previous day’s interactions as the signal, so the improvement loop stays tied to fresh production behavior.
Cons
- The public pages do not describe full pricing beyond the merged-PR model and a free first three PRs.
- The homepage names only a small set of integrations, so broader compatibility is not fully documented here.
- The workflow is centered on nightly analysis and one PR per morning, which may not fit teams that want immediate intervention.
FAQ
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.
Quick Facts
- 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
Альтернативы Basalt
blop
blop is a QA agent that writes browser tests as code in your repo, runs them in CI, clusters repeated failures, and can open PRs to fix broken tests.
Orca
Orca is an Agent Development Environment for shipping with coding agents, running multiple CLI agents in parallel across isolated worktrees, with desktop and mobile workflows.
Firebase Studio
Firebase Studio is a web-based workspace for full-stack app development with Gemini-assisted coding, app previews, cloud emulators, collaboration, and browser deployment.
AI Magicx
AI Magicx is a unified AI workspace for chat, image, video, voice, music, email and developer tasks, helping teams and creators manage multiple models in one place.
Paper
Paper is a design tool that connects canvas, code, and AI agents so teams can create, share, and ship work in one workflow. Includes desktop app and MCP access.
RLAMA
RLAMA is a local AI platform for building RAG systems and intelligent agents on macOS, Linux, and Windows, with HTTP API support.