Evidence beats fluency
A model that writes a confident paragraph about your outage is worse than useless — it is a liability with good grammar. The only claim worth shipping is one you can trace to a probe.
Tier2About
Evidence over vibes
Tier2 exists because tier-2 support is evidence work, and the industry answered that with chatbots that guess. We would rather ship an agent that shows what it checked and stops when it cannot.
Invite-onlyFive design partnersBuilt for technical support teams
The problem we kept watching
The expensive part of a ticket is the discovery
By the time an escalation reaches a senior engineer, someone has already spent forty-five minutes reproducing it, correlating it against a deploy, and reading a dashboard. That work is mechanical, it is well documented in your own runbooks, and it is being done by the most expensive people in the company.
What was missing was not intelligence. It was accountability: a way for an agent to do that discovery and then prove, line by line, what it actually observed. So we built the proof first and the agent second.
What we believe
Four opinions the product is built on
Drafts, not autonomy theatre
Nobody wants an agent emailing their largest customer at 2am. Approval is not a limitation we plan to remove; it is the reason the product is allowed near a customer at all.
The eng team is the renewal
Support leaders buy this, but engineers decide whether it survives the quarter. Every design call — dedup keys, read-only credentials, runbook diffs — is made for them.
Isolation is not a feature
One agent per tenant costs more to run than a pool. We pay it because an agent that accumulates memory across customers is a breach waiting for a date.
Invariants
Rules we will not trade for a feature
These are written into the codebase and enforced by hooks, not by good intentions. If a customer asks us to break one, the answer is no.
- Invariant
A claim without a probe_run_id is dropped
Enforced in the renderer, so it holds even when the model is having a bad day. It is the one rule the entire product rests on.
- Invariant
Read and probe only against your infrastructure
The agent executes freely in its own sandbox and not at all in yours. No write-capable credential is loaded, which is what makes the first half survivable.
- Invariant
Only reproduced reaches an engineer
The other four verdicts are drafted replies. Protecting that ratio is protecting the economics, so we treat it as a product constraint rather than a metric.
- Invariant
Corrections outrank everything the agent taught itself
An annotation from your engineer beats a fact learned from a ticket body, permanently. Provenance is recorded on every memory write for exactly this reason.
“If it cannot tell me which probe produced the claim, I am not going to forward it to a customer. That is the whole bar.”
Working with us
We are taking a small number of design partners.
Technical companies with a real tier-2 queue and engineers willing to correct an agent in public.
Draft-only by defaultRead + probe access only