We build the agent. You own it.
We scope it, build it, prove it against your own cases, and hand it over. It runs on your infrastructure, under your keys.
The gap is not the model.
Most agent projects die between the demo that impressed everyone and the first real customer. What runs out on that stretch is never intelligence.
It is a spec someone signed. Tools with hard limits. A suite allowed to block a release. A named owner after handover.
So we build where the work has a right answer, an audit trail, and a cost when it is wrong.
Five stations. One route.
Every build walks the same line, in the same order. No station is finished until it has left an artifact behind.
The suite runs on every change. A gate that fails stops the release.
One sample run on 240 cases, shown to explain the shape of a suite. Not a result you are being sold.
# Runs on every change. A gate that fails blocks the release.suite: ticket-resolvercases: 240 # from your own last quartergates: resolved_without_person: ">= 0.65" unsourced_answers: "<= 0.02" policy_breaches: "== 0" reply_seconds_p95: "<= 12" cost_per_case_usd: "<= 0.06"last_run: # a sample, not a result you are sold resolved_without_person: 0.71 # pass unsourced_answers: 0.012 # pass policy_breaches: 0 # pass reply_seconds_p95: 9.4 # pass cost_per_case_usd: 0.043 # passon_fail: block releaseWhat comes off the line.
Four shapes we build often, and the one thing each of them is never allowed to do.
- 01
Ticket resolver
never closes a case without a source it can show you
- 02
Onboarding operator
never creates a record until every field passes
- 03
Document extractor
never guesses a field, it escalates the page
- 04
Case router
never routes a case without recording why
Production is the deliverable.
A demo proves an agent can work once. We ship the one that keeps working.
- Traced
- Every run is recorded. You can open any decision the agent made and see what it read.
- Gated
- A change ships only when the eval suite, built from your own cases, passes.
- Checkpointed
- State is saved at every step, so a person can inspect, resume or correct a run.
- Yours
- It runs on your infrastructure, under your keys, and the repo is yours.
The first build is not the last.
Senda ships builds today. The two stages after it are the plan, and we say them as a plan.
- Start small
The build
One agent, on your infrastructure, owned by you.
- Grow
The catalogue
Shapes repeat inside one kind of operation. Your second agent of a shape starts with its spec, its graph and its evals already written.
- Scale
The fleet
The evidence that gates one release is the evidence you need to run ten. One place for what every agent did, cost and stopped on.
What people ask before they write.
- Why not build it in-house?
- Your team can build the demo. What we build is the spec, the limits, the eval suite and the runbook.
- What happens when the agent is wrong?
- The limits live in the tools, not in a prompt. Every run is traced, so a bad one can be replayed.
- What does a build cost?
- One price for one build, quoted with the first pass spec, before any work starts.
- What do we get at handover?
- The repo, the keys, the runbook, the traces and the eval dashboard, plus thirty days of fixes.
Tell us what to build.
Describe the job the agent should do. We reply within two working days with a first pass spec and a price.