The agent orchestration platform · Public beta

The durable runtime for production AI agents.

Design multi-agent workflows on a canvas, then run them on infrastructure that survives crashes, pauses for human approval, and puts a dollar cost on every step.

  • Design multi-agent workflows on a visual canvas
  • Every step checkpoints — runs survive crashes and resume
  • Human approvals, full audit, and a dollar cost on every step
Support Triage CrewRunning
run_af772b · v40 / 5 steps
  • Classify
  • Draft replyqueued
  • Fetch contextqueued
  • Human reviewqueued
  • Notify teamqueued
starting run…total $0.0000
10min
from empty terminal to a governed run
100%
of steps cost-attributed
99.9%
uptime target on Enterprise SLAs
2
SDKs — Python & TypeScript

Built for teams running agents in production

Illustrative placeholders — not current customers

VantageMeridianCorelineNorthbeamHelios SystemsStructaDatum LabsRidgeline
VantageMeridianCorelineNorthbeamHelios SystemsStructaDatum LabsRidgeline

How it works

Whiteboard to production in three moves.

01

Design

Compose agents, tools, conditions, and human gates on the canvas — or in code with the SDK. Graphs are validated before a single token is spent.

canvas · sdk · templates

02

Execute

Runs execute on a durable runtime that checkpoints after every node, with retries, timeouts, and bounded parallelism. Kill the process — the run resumes.

checkpoint · retry · resume

03

Govern

Approvals pause runs durably until a person decides. Every action lands in the audit log, and every step carries a dollar cost finance can read.

approve · audit · attribute

Platform

Everything between a prompt and production.

One platform to design, execute, observe, and govern multi-agent workflows — without stitching together five tools.

Classify
Fetch CRM
Human gate
draft · v4

Visual workflow builder

Design multi-agent systems on a canvas. Agents, tools, conditions, loops, and gates — each node compiles to real executable graph state.

ckpt_15 … ckpt_18resume point

state persisted after every node.
kill the process — the run continues.

Checkpointed execution

State is persisted after every node. Runs survive crashes and restarts, and resume from the last checkpoint.

awaiting approval

Review the drafted refund of $1,240 before it is sent.

ApproveReject

Human-in-the-loop

Gate any step behind a structured approval — multi-stage, conditional, with reminder and escalation ladders. The run pauses, durably, until a person decides.

classify
$0.0004
draft
$0.0031
review
$0.0000
notify
$0.0001
total / run$0.0036

Cost attribution & budgets

Every LLM call is metered — cost per step, per run, per outcome. Cap spend with budgets enforced on every trigger, with atomic reservations and audited overrides.

approverun af772bl.friedman
updatewf 9bc62fapi-key-2
runwf ceb7afscheduler
createwf cdcb40l.friedman

Audit & policy events

Every change, run, and approval is logged — and every runtime policy decision (denials, redactions, blocks, overrides) streams as an inspectable event. Compliance is built in, not bolted on.

claude-sonnet-4-6

primary

claude-haiku-4-5

fallback 1

gpt-4o-mini

fallback 2

Multi-LLM routing

Anthropic and OpenAI behind one interface, with per-node model selection and governed fallback chains — failover only for approved error classes, never to a model your policy forbids.

REST APIPython SDKTypeScript SDKWebhooksMCP serverMCP clientSlackCustom tools

Connect anything

MCP in both directions, webhooks, a full REST API, and typed SDKs. If it speaks HTTP, it plugs into the graph.

Reliability

Works in the demo.
And in production.

Most agent projects die at the same wall: they work once, then break under real load. Ballast is built for the second week, not the first demo.

  • Automatic retries

    Transient failures are retried with exponential backoff — per node, with configurable limits.

  • Per-step timeouts

    No agent hangs a workflow. Every node has a hard timeout and a graceful failure path.

  • Resume from any checkpoint

    Kill the process mid-run. State is already in the database, and the run picks up exactly where it stopped.

ballast — production
We built Ballast to replace a pile of cron jobs and retry logic with one governed engine — so runs survive your deploys, refunds wait for a human, and finance finally sees a dollar cost per outcome.

The Ballast team

Why we built it

Observability

See every step. Cost every call.

Full trace of every run — model calls, tool invocations, approvals — with duration and dollar cost attributed per step. Debugging a multi-agent failure stops being archaeology.

run r_7f3a · support-triage · v4succeeded

classify_ticket

agent

312ms$0.0004

fetch_context

tool

587ms$0.0000

draft_reply

agent

1.9s$0.0031

human_review

gate

4m 12s$0.0000

notify_team

tool

142ms$0.0000
cost per successful outcome$0.0035

For engineers

Operators get a canvas.
You get an API.

Everything on the canvas is a typed SDK call away. Trigger governed runs from your product, resolve approval gates programmatically, and register your own agents as tools.

  • Python & TypeScript SDKs, one-for-one with the REST API
  • Idempotent triggers, run polling, and typed errors built in
  • MCP in both directions — expose workflows as tools, or call external servers mid-run
Read the SDK docs
from agentos_sdk import AgentOS client = AgentOS("https://api.ballastos.com", api_key="aos_...") run = client.run(    "Support Triage Crew",    {"ticket": "I was double charged"},    wait=True,                      # poll until terminal    idempotency_key="ticket-8841",  # safe to retry) if run["status"] == "paused":       # held at a human gate    client.approve(run["id"], input={"reviewer": "lukas"}) print(run["status"], run["output"])

Enterprise-ready

Governance your security team signs off on.

Audit, access control, and human approval are core engine features — not an enterprise upsell bolted on later. The controls regulated teams ask for are here on day one.

SOC 2 Type II

Controls mapped to SOC 2. The Type II audit is planned — on our roadmap, not yet started.

SSO (OIDC)

Bring Okta, Entra, or Google. Just-in-time provisioning with enforced default roles.

SCIM 2.0

Automatic user provisioning and deprovisioning driven by your identity provider.

MFA & lockout

TOTP two-factor for every member, with brute-force login lockout on by default.

Role-based access

Ten granular capabilities with per-member overrides and role-scoped API keys.

Immutable audit log

Every create, run, approval, and rejection recorded — exportable as CSV.

Encryption & isolation

TLS in transit, secrets encrypted at rest, strict per-workspace data isolation.

Self-host / VPC

Run the same container in your own cloud — no run data leaves your perimeter.

Pricing

Pay for runs, not seats.

Usage-based pricing that scales with your agents. Hard caps by default — no surprise bills.

Free

$0forever

For prototypes and solo builders.

  • 1,000 runs / month
  • 5 workflows
  • Visual builder
  • 7-day trace retention
  • Community support
Start free

Starter

$49/month

For small teams shipping their first production agents.

  • 10,000 runs / month
  • Unlimited workflows
  • Checkpoint resume
  • 30-day trace retention
  • Email support
Start with Starter
Most popular

Pro

$299/month

For teams running agent fleets that matter.

  • 200,000 runs / month
  • Governance & audit log
  • Human approval gates
  • 90-day trace retention
  • Team collaboration
  • Priority support
Start with Pro

Enterprise

Custom

For regulated industries and large orgs.

  • Custom volume & SLAs
  • SSO (OIDC) & SCIM
  • IP allowlist
  • Unlimited audit retention
  • Self-host option
  • Dedicated support
Talk to sales

LLM provider costs pass through at cost on all paid tiers. Overage is cap-and-notify by default.

FAQ

Questions, answered.

Start free — no API keys

Ship agents you can trust in production.

Design your first checkpointed workflow, run it on the mock model, and watch every step and its cost — in under thirty minutes.