Driftlane is the trace-first observability platform for distributed AI agents. Capture every prompt, replay every failure, and ship to production without bolting on five separate dashboards.
No credit card. SDK in 10 lines. Free forever for solo developers.
Three failure modes show up in every production agent stack — and traditional APM tools were never built to catch them.
Your agent returned a "completed" status — but the answer was wrong, the tool call was skipped, or the user got a polite hallucination. Nothing throws an exception. Your alerting stays green. The complaint email shows up four days later.
A single ten-cent agent run quietly turns into a four-dollar loop. Your monthly LLM spend is a black box. By the time finance asks why the bill 4x'd, you cannot point to a single trace, customer, or prompt change as the cause.
Your evals passed in CI. Real traffic does things your fixtures never imagined. Without a way to replay actual production traces against new prompts or models, every release is a coin flip and every regression is a fire drill.
One SDK, one data model, three primitives. Everything else — costs, evals, alerts — falls out of the trace.
One import wraps every LLM call, tool call, and retrieval. Driftlane records inputs, outputs, latency, and cost — with PII redaction before bytes leave your process.
Spans roll up into a single tree per agent run. See the full reasoning path, branching decisions, retries, and which tool returned the bad payload — searchable across millions of runs.
Pick any production trace and re-run it against a new prompt, a new model, or a patched tool. Diff outputs side by side, score with evals, and promote the winner — no synthetic dataset needed.
What a normal day looks like before Driftlane — and what it looks like once your agents are instrumented.
Six primitives that replace the half-dozen tools you would otherwise stitch together.
OpenTelemetry-compatible spans roll up into a single tree per run. Search, filter, and pivot across millions of agent executions in under a second.
Slice token spend by agent, customer, feature flag, or release tag. Get alerted before a runaway loop burns through your monthly budget.
Re-run any production trace against a new prompt, model, or tool. Diff the outputs, score with evals, and promote the winner without a synthetic dataset.
Helpfulness, faithfulness, toxicity, JSON-validity — ship with a battery of LLM-as-judge scorers, or plug in your own Python function as a custom eval.
Mask emails, names, and custom regex patterns inside the SDK before payloads leave your process. Self-host the control plane for VPC-isolated workloads.
Slack, PagerDuty, and webhook alerts when latency p95, error rate, or spend per tenant crosses your thresholds. Tied directly to the offending trace.
Pay for the spans you ingest. Invite the whole team for free. Cancel any time.
For solo developers shipping their first agent. No card required.
Get startedFor small teams putting agents in front of real customers.
Start free trialFor teams running agents at production volume across multiple environments.
Book a demoBuilders running real agents in production — not demos.
"We used to chase agent bugs across CloudWatch, Sentry, and three Postgres tables. Now it is one trace. My on-call engineers actually sleep again."
"Within seven days of instrumenting our support agent, Driftlane surfaced two tool regressions and a silent rate-limit retry storm. The replay feature paid for itself before the trial ended."
"Cost attribution by tenant let us spot two enterprise accounts looping on a bad retrieval prompt. One config change cut our OpenAI bill almost in half."
"Replaying yesterday's traffic against a new prompt before merging completely changed our release cadence. We went from one prompt change a week to four a day, with fewer regressions."
"SDK in, env var set, first trace on the dashboard before my coffee finished brewing. The dev experience is the closest thing to Stripe I have seen in AI tooling."
"We process millions of agent runs a month. The Driftlane SDK never blocks our main loop, and the disk-buffered replay saved us during a noisy AWS week. Zero dropped traces."
I spent four years at Stripe building the observability stack behind Radar, then two years at Anthropic on the team that wrote the first internal tooling for agent evals. Driftlane is the product I kept wanting at both companies and could never find on the market.
We are a small, distributed engineering team — most of us have shipped production AI systems before starting Driftlane. We backed by Y Combinator (F25), General Catalyst, and a handful of operators from OpenAI, Vercel, and Datadog.
If Driftlane does not earn its place in your stack by day 14, your account quietly downgrades to the free Solo tier. No clawback emails, no friction.
No. The Driftlane SDK is a single import that wraps your existing LLM and tool calls. Most teams instrument their first agent in under ten minutes — typically two lines in your entrypoint plus an environment variable for your project key. You keep your framework of choice, whether that is LangGraph, CrewAI, Vercel AI SDK, or hand-rolled.
Out of the box we trace OpenAI, Anthropic, Google, Mistral, Groq, Together, and any provider that follows the OpenAI-compatible chat schema. On the framework side we have first-class adapters for LangChain, LangGraph, LlamaIndex, CrewAI, Vercel AI SDK, Pydantic AI, and the Anthropic Agent SDK. Anything else can be instrumented with our generic span API.
Redaction happens in the SDK before the payload ever leaves your machine. You can configure regex masks, key allowlists, or run our local PII model. For regulated workloads we offer a self-hosted control plane that keeps trace bodies inside your VPC while metadata streams to our cloud for the dashboard view.
Tracing is fully async and batched. The SDK adds under one millisecond per span on a warm process and never blocks your main loop — even on cold start the overhead stays below five milliseconds. If our ingest is unreachable, spans buffer to disk and replay automatically once connectivity returns.
Yes. Every captured trace is replayable. Pick any subset — failed traces, expensive traces, traces from a specific customer — and re-run them against a new prompt, a new model, or an updated tool. Diff the outputs side by side, score with built-in or custom evals, and promote the winning configuration.
Team is a flat forty-nine dollars per month and includes one hundred thousand spans plus seven-day retention. Scale is four hundred ninety-nine per month for two million spans and thirty-day retention. Overage is metered per ten thousand spans and never surprises you — we pause ingest at your configured cap and email you first.
Scale tier includes a self-hosted control plane image you can deploy into your own Kubernetes cluster. It ships as a single Helm chart with Postgres and ClickHouse backends. Enterprise customers get a hardened build with SSO, audit logs, and BYOK encryption — happy to walk you through it on a call.
Teams building production agents — not chatbots. If your system makes multiple LLM calls, hits external tools, branches on results, and needs to be debuggable at 2 a.m., Driftlane is built for you. Today our customers run customer-support agents, coding agents, research agents, and internal ops automations at companies from two-person YC startups to publicly traded fintechs.
Free Solo tier, no card, no sales call. Upgrade when your traffic does.