Your agent → your collector → one pipeline7 products1 Private betaStation SouthBase
Know why your agent failed — and trust the measurement that said it did. Tracing is a solved problem; OpenTelemetry commoditised it. The unsolved problems are whether an agent's output was actually good, and — when it wasn't — which step caused it.
SouthBase is pre-launch. The platform is in private beta with design partners and there is no self-signup yet. Every product page states exactly where that product stands.
Speaks OpenTelemetryYOUR TRACES ARE YOURS
Your application, your collector, one pipeline. Redaction happens before anything crosses the boundary — which is the only promise about your data we can make structurally rather than contractually.
Emits OpenTelemetry spans through an OpenInference instrumentor. No proprietary SDK.
OTLPAn Apache-2.0 OpenTelemetry Collector distribution running in your own infrastructure. Redacts fields, normalises tokens into four buckets, attaches cost.
Apache-2.0Spans land blob-first, then a worker computes cost against a versioned price book and writes the result with the version that produced it.
OTLP/HTTPOne pipeline feeds observability, evaluation, governance and the security products — rather than four agents and four bills.
one pipelineThe loopTHREE JOBS · ONE PIPELINE
Everyone converged on tracing. Almost no one solved trustworthy evaluation or failure attribution — and those are the two that tell you what to fix. Here is where each of the three actually stands.
01
A trace that survives a ten-minute run
Span trees are the right shape for a request and the wrong shape for an agent run with fifteen tool calls and sub-agents. The transcript reads in order, start to finish, with the hierarchy a toggle away.
02
A judge you can calibrate, not just more metrics
LLM-as-judge is widely distrusted for documented reasons — position bias, verbosity bias, self-enhancement bias, poor correlation with human raters. More metrics do not fix that. Measuring the judge against human labels and publishing the interval does.
03
Which step actually caused the failure
When a multi-agent run fails, the best tools give you a trace and one black-box score. They cannot say whether planning, retrieval, synthesis, a tool call or a memory write caused it. The literature calls credit assignment underdeveloped, and it is.
Decomposed the request into four sub-tasks.
Returned a passage that no longer matched the source document.
Reasoned correctly over the passage it was given.
Produced a confident answer grounded in the stale passage.
The platform7 PRODUCTS
The same instrumented traces answer four different kinds of question — did it work, was it good, was it allowed, and is it secure. Building those on one pipeline is the point; it is also why the list is longer than what ships today.
Know what your agents did, whether the output was good, and which step broke when it wasn't.
The controls and evidence that let an AI system pass review — over the same pipeline.
What actually works today1 Private beta
A roadmap is not a product. Here is the part that exists and runs, so you can judge us on it rather than on the list above.
Input, output, cache creation and cache read are billed at four different rates. A tool reporting two cannot tell you what a request cost. Take a real shape from our conformance fixtures — 500 fresh input tokens, 220 output, 40,000 cache reads. A two-bucket tool reports “40,500 input tokens”, which is arithmetically true and completely useless.
cache_read / (input + cache_creation + cache_read) = 0.988
That is the cache-hit ratio, and it is the number that tells you prompt caching is working. How the four buckets work →
Span trees are the right shape for a request and the wrong shape for a ten-minute agent run with fifteen tool calls and sub-agents. The transcript reads in order, start to finish, with a span-tree toggle when you want the hierarchy back.
Capture enough of a run — prompt, params, model, tool inputs and outputs, retrieved context — to re-run it and diff the result against the original. It is the most frequently requested feature in this category and one of the simplest to actually use.
The collector is Apache-2.0 and runs in your infrastructure. Redaction happens there, so sensitive payloads never reach us — and you can read the code that does it rather than take our word for it.
Against the alternativesFIGURES VERIFIED 2026-08-08
Every comparison page carries at least one row where the other tool wins, and a plain statement of who should stay with them. A comparison with no concessions is marketing, and readers discount all of it.
On lock-in
The wire format is OpenTelemetry, so the instrumentation that points at SouthBase points at any other OTLP backend. That is deliberate, and it is worth being precise about why: we are a hosted, proprietary platform, so “your data is portable” is a promise we should have to demonstrate rather than assert. Using a neutral wire format is how it gets demonstrated.
Go deeperNO GATED PDFS
There is no whitepaper and no webinar. The specification, the arithmetic and the source are the marketing.
The quickest way to judge this is arithmetic: run a cache-heavy Claude workload through it and see whether the number matches your Anthropic console. That is the bar we set for ourselves — and design partners are how it gets tested against real traffic.