Hosted agents

Deploy an agent that
remembers yesterday.

One click deploys a persistent agent with its own identity, durable memory, and an inbox. It wakes when a message arrives, works under a metered budget, and goes back to costing nothing.

the whole onboarding
curl https://api.swarms.dev/api/v1/agents \  -H "Authorization: Bearer $SWARMS_KEY" \  -d '{    "template": "hermes",    "name": "ops-analyst",    "wakeBudgetUsd": 0.25  }' # → 201 { "agentId": "agt_9d4…", "status": "idle" }# Hired in one call. Asleep until needed.

Most agents are goldfish brilliant for one request, blank by the next. A hosted agent has a name, a memory, and an inbox. It recalls Tuesday’s decision on Wednesday, wakes the moment a message lands, and costs almost nothing in between. That is not a function call. That is staff.

Durable memory

It remembers Tuesday on Wednesday.

Stateless agents make you the memory: every session starts with you re-explaining the project. A hosted agent carries a persistent identity and durable memory stored in Postgres — decisions, preferences, and open threads survive every restart.

The interface is a message, not a prompt. POST to its inbox and the reply builds on everything that came before, the way the third week with a colleague beats the first day.

talking to a colleague, not a session
POST /api/v1/agents/agt_9d4…/messages{  "content": "Same report as yesterday,              but add the churn numbers."} # "Same as yesterday" just works — yesterday lives# in Postgres, not in a context window that# expired at midnight.

Wakes, works, sleeps

Always on call. Almost never on the clock.

The agent wakes on two triggers: a message in its inbox, or a heartbeat you configure. Between wakes it holds no compute — an idle agent costs approximately nothing, the way an employee off shift costs no overtime.

Every wake is a real run with real controls: metered to the GPU-second and capped by a per-wake budget. A stuck task stops at its ceiling instead of burning through the night.

a persistent agent's afternoon

idle — costing nothing

→ “what changed in our competitors' pricing this week?”

Employer controls

Pause it, resume it, read its payroll.

You stay the employer. Pause, resume, or terminate any agent in one call — a paused agent keeps its memory and answers nothing until you say otherwise.

And the payroll is itemized: every dollar is attributed to the specific agent that spent it, wake by wake. When one costs more than it returns, the numbers say so before you have to guess.

schedule: "0 6 * * *" — the morning report

Mon 06:00

$0.14

Tue 06:00

$0.14

Wed 06:00

$0.14

Thu 06:00

$0.14

Stop renting “a model you re-brief every session.” Start employing “a colleague who read yesterday’s thread and answers before you are back at your desk.”
The mental model

What that unlocks

Standing roles, staffed

An ops analyst that owns the morning report. A triage agent that owns the support queue. Deployed once, on duty indefinitely.

Context that compounds

Week three beats week one because nothing was forgotten in between. The agent gets more useful the longer it holds the job.

Idle costs that round to zero

You pay for wakes, not uptime. An agent that answers forty messages a month bills like forty small runs — not like a server.

Accountability per agent

Per-agent spend attribution turns “what are the agents costing us” into a table with names on it — and a pause button next to each.

Keep exploring

Give your agent a workforce.

Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.