AgentOS is observability and control for AI agents at work: every run streamed live, every tool call, token, and cent on the record, and a hard pause for your approval before anything sensitive happens.
Works with any language or runtime: Python, Node.js, Go, or plain curl
The problem
The scripts work in a demo. Then they meet the real company: customers, money, production data. Three things break at once:
// NO VISIBILITY
Agents scattered across laptops, cron jobs and notebooks. When one misbehaves, there's no record of what ran, what it cost, or why it failed.
// NO CONTROL
An autonomous agent that can act is an agent that can act wrongly. Without approval gates, the only safe agent is a useless one.
// WRONG PRICING
AI vendors charge for every human seat. But the whole point of agents is that software does the work. Paying per person taxes exactly the adoption you want.
Observability
This is the part nobody else takes seriously. AI agent observability isn't a log file. It's knowing, for every run of every agent, exactly what happened, what it cost, and why. AgentOS records it all by default.
// LIVE
Every run streams into the dashboard event by event: the prompt, each tool call, each model response, in real time. When an agent is working, you can literally watch it think.
// RECORDED
Full run history with millisecond timestamps: inputs, outputs, tool calls, tokens, cost per run, and every human approval decision. When someone asks “what did the agent do?”, the answer is one click, not an investigation.
// WATCHED
Alerts fire on failures, silence, or slow runs. Built-in intelligence agents audit the fleet for regressions, anomalies, token waste, and weakening prompts, and file findings before you'd have noticed.
And it works for agents you already have: three SDK lines or an OpenTelemetry export puts any agent, in any language, on the record. See how instrumentation works →
Proof, not promises
These aren't demos. They're the actual agents operating AgentOS, the company, on AgentOS, the product, right now. Our inbox, our support queue, our outreach, our social feed, our on-call briefings: agents do the work, humans approve what matters.
Reads every inbound email, classifies it (support, sales, billing, spam) and files it with the right label. Genuinely ambiguous mail stays in the inbox for a human.
Drafts grounded replies to support email and sends the routine ones. Complex or sensitive threads become drafts for human review instead of auto-sending.
Sends personalised outreach from the prospect sheet, tracks what was sent, and surfaces replies for a human to handle.
Drafts a week of LinkedIn and X posts every Monday and emails them for review. Publishing is approval-gated: nothing goes live without a human clicking approve.
Fetches competitor blogs and changelogs, summarises what each one shipped last week, and emails a structured intel brief.
Summarises overnight alerts and incidents into a single what-happened-what's-open brief for whoever is on call.
Want agents like these? Install ready-made agents from the marketplace in one click, then connect your tools.
Features
Every event, tool call, and LLM response captured with millisecond precision. When an agent did something, you can see exactly what and why.
Conditions on error rate, duration, silence, or custom events. In-app, email, or webhook. You hear about problems before your customers do.
Built-in intelligence agents watch your agents for anomalies, regressions, prompt weaknesses, and token waste. Daily AI summaries keep you ahead of it.
Gate any tool behind sign-off. The agent pauses mid-run, your team gets notified, and nothing sends, posts, or spends until a human approves.
Define the job, tools, and model in the dashboard. AgentOS runs the whole loop: no infra, no cold starts, no ops. Invoke with one HTTP call or a schedule.
Install proven agents in one click, including the ones running our own company. Skip the blank page; start from something that already works.
Already have agents in code? Instrument them in three lines (startRun, emit, complete) and they show up in the same dashboard, approvals and all.
POST events from any language, or point your OTel GenAI traces at our OTLP endpoint. If it runs anywhere, AgentOS can supervise it.
Connect Claude, Cursor, or any MCP client to your workspace. Manage agents, approve runs, and read reports without leaving your editor.
For developers
Already built agents? Keep them where they are: add three SDK lines and they get the same run history, approvals, and alerts as hosted agents.
Pricing
Cheaper than the work it replaces. Start free, no credit card.
Put your first agent to work.
For teams handing real work to agents.
or $390/year billed annually (2 months free)
Start with TeamFor companies running agents at scale.
AI usage is bring-your-own-key: agents run on your OpenAI, Anthropic, or any compatible endpoint, billed by the provider at their normal price. AgentOS takes no markup on your tokens, ever.
FAQ
No. Install an agent from the marketplace, connect your tools with OAuth, and it starts working. Writing your own agent is describing the job in plain language: the prompt, the tools it may use, and which of those need your approval. Developers get SDKs and APIs on top, but they're optional.
That nothing an agent does is invisible. Every run is streamed live and stored in full: each tool call, each model response, tokens, cost, and any human approval decision, with millisecond timestamps. You can watch a run as it happens, replay any past run step by step, and set alerts on failures, silence, or slow runs. It works for hosted agents automatically, and for agents you already run elsewhere via the SDK or OpenTelemetry.
Not if you gate it. Any tool can be marked as requiring approval, whether it sends email, posts to social, or writes to a spreadsheet. A gated call pauses the run, lands in your approvals inbox, and nothing happens until a human approves or rejects it. We run our own social publishing this way.
AgentOS hosts and runs your agents: the LLM calls, the tool executions, the schedules, the retries. There's nothing for you to deploy or keep alive. If you'd rather run agents in your own code, the SDK reports them into the same dashboard.
No. Add the SDK (npm or pip) and wrap your code with startRun/complete. It takes about 10 minutes. Or export OpenTelemetry traces to our OTLP endpoint with zero AgentOS-specific code. Your agent keeps running wherever it is.
Yes. Run data is scoped to your workspace and only accessible by workspace members. Credentials for connected tools are encrypted at rest, and secrets are readable only by workspace admins. We do not use your data to train models. An on-prem / self-hosted option exists for Enterprise.
Hosted agents run on Anthropic and OpenAI models out of the box, plus any OpenAI-compatible endpoint (including self-hosted models on Enterprise). Agents you run through the SDK can use anything. AgentOS supervises the run either way.
Different job. ChatGPT is people chatting with AI, and it's good at that. AgentOS is software doing jobs on its own: scheduled, supervised, auditable, with a human gate on anything sensitive. Many teams keep both: seats for the humans, AgentOS for the workforce of agents. What we'd argue against is paying per-person prices to run software.
$0 and an afternoon. The free plan has 2 agents, unlimited teammates and no credit card. Bring one annoying, repetitive workflow (reporting, triage, chasing, monitoring) and give it to an agent with an approval gate on the scary part. That first supervised run is usually the whole sales pitch.
Install one of the agents that runs our company, gate the scary part behind your approval, and watch the first supervised run. If that doesn't convince you, nothing we write here will.