company
We build the boring parts of AI agents, on purpose.
Retries, approvals, budgets, tests, and traces. None of it makes a demo look better. All of it decides whether an agent survives its first month in production. That is the work we chose.
our story
Started by two engineers who got paged at 3 a.m.
Before Trellis, our founders ran support and operations agents at their old jobs. The models were good enough. Everything around them was not. Runs died halfway through a refund. A prompt change broke a workflow nobody was watching. A loop burned a month of budget over a weekend.
In 2023 they started building the runtime they had wanted: runs that pick up where they stopped, rules the model cannot argue with, tests that block a bad change, and a trace that shows every step and its price.
Today Trellis is a small company with a narrow job. We do not build models or chat apps. We build the layer that lets your engineers trust the agents they ship.
The company, in four numbers
sample figures · fictional company
- founded in San Francisco
- 2023
- people on the team
- 64
- time zones covered by on-call
- 4
- minimum notice before a breaking change
- 90 days
values
Five rules we hold ourselves to.
Each one shows up somewhere you can check: the rate card, the changelog, the essays, or the product itself.
01 · show-the-math
Show the math
Every price, metric, and estimate on our site comes with its formula or its measurement window. If you can't check it, we shouldn't print it.
in practice: Our rate card and configurator use the same code as our invoices.
02 · say-when-not-to-buy
Say when not to buy
Most workflows are better as a simple pipeline. We tell teams that in the first call, even when it costs us the deal.
in practice: Our CTO wrote an essay on when not to use an agent, and we link it from the homepage.
03 · budgets-are-law
Budgets are law
A limit that the model can ignore is a suggestion. We enforce caps, schemas, and allowlists in the runtime, where prompts can't reach them.
in practice: The step that would break a budget never runs.
04 · traces-over-opinions
Traces over opinions
When we argue about agent behavior, internally or with a customer, we open a trace. Data ends the meeting faster than seniority does.
in practice: Support tickets start with a run ID, not a screenshot.
05 · boring-on-purpose
Boring on purpose
Retries, checkpoints, and audit logs are not exciting. They are what lets an agent run at 3 a.m. without waking anyone up.
in practice: Breaking changes get a migration guide and at least 90 days' notice.
read the proof
"When not to use an agent," by our CTO →
history
Four years, one narrow job.
Every module started as a customer asking for the same fix twice. The full release record lives in the changelog.
-
2023
Two engineers who had run agent fleets at past jobs start Trellis to fix the parts that kept breaking: retries, approvals, and runaway spend.
-
2024
The durable runtime and guardrails leave private beta. The free dev tier opens with 1,000 runs a month.
-
2025
Evals join the platform after customers asked how to test a model swap before it shipped.
-
2026
Runs API v2 aligns traces with OpenTelemetry. Eval gates land in CI and guardrails get a dry-run mode.
team
Every module has a name on it.
Like a CODEOWNERS file: each lead owns a part of the product and answers for it, including on customer calls. The people below are fictional, shown with generated avatars.
-
Elena Marsh
Co-founder and CEO
owns: customers, pricing, the rate card
Ran platform teams for ten years before co-founding Trellis. Reads a customer trace review every week.
-
Sam Okafor
Co-founder and CTO
owns: runtime/, checkpoints, the scheduler
Built the first durable runtime in a garage-sized office. Wrote the essay on when not to use an agent.
-
Noor Haddad
Founding engineer
owns: evals/, graders, CI gates
Turned the team's eval habit into the Evals module. Believes every failed case should be kept forever.
-
Tomás Reyes
Guardrails lead
owns: policies/, screens, enforcement
Designed the observe, dry-run, and enforce rollout path after watching one too many policies ship straight to enforce.
-
Aiko Tanaka
Observability lead
owns: traces/, rollups, OTel export
Makes sure every span carries its tokens and its price, and that customers can take their data with them.
-
Ruth Adeyemi
Head of solutions engineering
owns: demos, migrations, on-call for customers
Leads the working sessions. Will tell you on the first call if a pipeline would serve you better.
how we work
Remote-first, with a desk in San Francisco.
Four time zones of on-call
Most of the team works remotely. On-call rotates across four time zones, so a sev-1 page reaches an engineer who is awake.
Everything public that can be
The rate card, the changelog, the status page, and the reasons we might not be a fit are all on this site. If you have to ask sales for it, we probably did it wrong.
Engineers on every call
Demos, support, and migrations are run by engineers who build the product. No hand-offs, no scripts. Contact the team.
$ trellis estimate --honest
Meet the people who would run your demo.
Price your stack in the configurator first, or book a working session with the engineers who own each module.