console preview — illustrative
AgentDeck
Ship production AI agents on top of your own data
AgentDeck is the scaffolding teams keep rebuilding around LLM agents — typed tools, retrieval on your indices, evals, guardrails, tracing — packaged so you ship the agent, not the plumbing. Model-agnostic, wired for the data you already stream.
Design-partner phase — a handful of teams, opening soon
01 — Problem
Why this exists
The demo agent takes a weekend; the production agent takes a quarter. Between the two sit tool contracts, retrieval freshness, eval suites, guardrails, cost control and tracing — infrastructure every team rebuilds from scratch, under deadline. None of it is your product.
02 — Capabilities
What it does today
Typed tool orchestration
Tools are typed contracts with retries, timeouts and cost budgets built in — the agent can only do what you declared.
RAG on live data
Retrieval wired to Elasticsearch and OpenSearch, refreshed by your Kafka streams — answers built on data from now, not from the last batch job.
Evals as CI gates
Eval suites built from recorded traces run on every change, in GitHub Actions or your CI of choice, and block regressions before deploy.
Guardrails & cost ceilings
Input and output policies, tool allow-lists and hard spend limits per run — a misbehaving run fails closed instead of surprising you on the invoice.
Full-loop tracing
Every plan, tool call and token accounted for in OpenTelemetry traces — debug an agent the way you debug a service.
Model-agnostic by design
Bring your own LLM API key — the tool, eval and tracing layers work the same whatever the provider.
03 — How it works
From connection to production
- 01
Declare tools & data
Describe your tools as typed contracts and point retrieval at your Elasticsearch or OpenSearch indices — kept fresh by the streams you already run.
- 02
Compose the loop
Assemble plan–act–observe agents from those tools, with retries, timeouts, budgets and guardrails as configuration instead of custom code.
- 03
Gate with evals
Record real traces, turn them into eval suites, and run them in CI. A change that regresses the agent never reaches production.
- 04
Ship & observe
Deploy behind a simple API. Every step of every run is traced and exported through OpenTelemetry, next to the rest of your stack.
04 — Use cases
Where it fits
Engineering teams that want AI agents in production without rebuilding the plumbing.
Eng lead shipping an internal copilot
A support or ops assistant over internal docs and tickets. AgentDeck supplies retrieval, guardrails and evals — your team's code is the workflow, not the framework.
Platform team automating runbooks
An agent that watches your streams, drafts the diagnosis and executes the safe steps of the runbook — every action typed, budgeted and traced.
The prototype that must become a product
The notebook agent works; production is another job. Keep your prompts and logic, adopt tools, evals and tracing from AgentDeck instead of building them.
FAQ
Product FAQ
Which models does it support?
AgentDeck is model-agnostic — you bring your own LLM API key, and the tool, eval and tracing layers work with any provider.
SaaS or self-hosted?
Designed to run in your own cloud, next to your data and your keys. That's the deployment we're validating with design partners; a managed option may come later.
How do evals actually run?
You record real agent traces, curate them into scenario suites, and AgentDeck replays them in CI on every change — scoring outputs and failing the pipeline on regression.
What does “design partner” mean?
A handful of teams building real agents with us before launch: hands-on onboarding, weekly iterations, and a real say in what ships first.
Build it with us
A small group of design partners gets direct access to the founder and real weight on what ships first. If the problem above is yours, let's talk.