The agent studio

Describe an agent. We handle the details.

Calafia builds your agent from components optimized for your specific task — model, memory, sources, schedule, spend cap. Run it here, or eject the code and take it anywhere.

You approve the full proposal before anything runs — sources, schedule, spend cap, all visible up front.

Watch a competitor's pricingMonitor a status pageCatch a real market moveA weekly arXiv brief

Your sentence comes with you through sign-in — describe first, create the free account when the proposal is worth keeping.

Market Move Alert
From the example sentence: “Track a stock and email me a brief when it moves more than 5% in a day”
Example proposal
Sources
Market quotes — public data (Yahoo; ~15-min delayed)
Email delivery — to your inbox
Schedule
Checks every 30 minutes — the watch default; say “every 5 minutes” or “daily at 8am” to change it
Spend cap
$5.00/day — hard cap; every check logged
Model
Gemini 2.5 Flash
Delivery
Emails only on a real move — silent otherwise
Describe your own agent to get a live proposal in its place.
Live from the registries

Counted, not claimed

These numbers are computed from this deployment's registries when the page renders — the same lists the builder composes from. Each tile links to the page that itemizes the lot.

Real sends · last 72 hours

What agents here delivered, lately

Dated sends from live agents on this deployment — briefs, watches, digests. Never a mockup; a quiet stretch renders nothing here.

Eng Morning Brief22 hours ago
Eng: 16 items need you
Founder Daily Brief23 hours ago
Brief: No New Items (2026-09-19)
My-Stuff Recall Alert1 day ago
New recall for Peanut Butter products — September 18, 2026
How it works

Describe → propose → approve → run → tune

01
Describe it
One or two sentences in the box above. The builder checks what it can actually verify before promising anything.
02
Read the proposal
It comes back as a card you can read in full: the sources it will use, the schedule it will keep, the budget it may spend.
03
Approve it
Nothing runs until you say so. Edit the sentence and get a new proposal, or approve this one as-is.
04
It runs on schedule
Checks on its own cadence, spending-capped, with a log of every check. Results land in your inbox.
05
Tune it in the agent lab
Test different combinations of your agent's core components — memory, model, planning, guardrails, delivery — and keep the one that wins. Hyperparameter tuning for agents.
The composition axes

Ten parts, all swappable

Every agent here is a composition. What it remembers, which model runs it, how it plans, what it may do — each is a setting you change on a running agent, not a rewrite. Each page below teaches one axis and lists exactly what this deployment carries, live from its registries.

MemoryEvaluationPlanningRetrievalStateGuardrailsModelsFrameworksToolsDelivery

See every part this deployment carries →

Built with it

What this machinery already runs

Ten swappable parts
Memory, models, planning, guardrails, delivery — every axis of a running agent is a page that teaches the trade-off and lists exactly what this deployment carries, live from its registries.
See the composition system →
The test bench
Same task, N compositions, one fixed eval — the experiment runner behind every agent's Tuning page. Change one part, read the next run, keep what worked.
How evaluation works →
The blog
Write-ups from running these agents in production — what worked, what failed, with the numbers from the logs.
Read the findings →

Products built on this platform run on their own sites.

Describe your first agent.

One sentence. You approve the proposal before anything runs.

Start describing

From the founder of Agentic Control Plane — every agent is spending-capped, keeps a log of every check, and stops with one click.