Starter
Get one agent working for you — self-serve, single-agent.
$49/mo
- Up to 3 agents
- 500 tasks/mo
- Self-healing retries
- Batch submission + webhooks
- Memory (success/failure patterns)
Autonomous Agent Operations API
Create an agent. Give it a task. It gets it done — or tells you honestly why not.
Create an agent, give it a task, and it gets it done — or tells you honestly why not.
Create named AI agents with a role, goal, and permitted tools. Send them tasks over a REST API. Agents plan the work, execute it, diagnose and recover from failures with hard retry limits, and remember what worked — without ever claiming success that hasn't been verified. Supervisor agents can delegate to specialist agents and verify their results before returning a final answer.
What it does
Create agents
Name an agent, give it a role, a goal, and instructions. Assign the tools and permissions it's allowed to use.
Send it tasks
Submit work over a REST API. Poll for status through queued, planning, working, retrying, and completed/failed.
Self-healing
A failed step is diagnosed and retried with a genuinely different approach — never the same attempt twice, and never an infinite loop.
Memory
Agents remember what worked and what didn't, isolated per customer account, and reuse successful strategies on future tasks.
Multi-agent orchestration
A supervisor agent can delegate to specialist agents and verifies their results before responding.
Bring your own tools
Connect any HTTP API or webhook — your CRM, Slack, GitHub, or internal systems — without waiting on us to build an integration.
No mandatory polling
Pass a webhookUrl and get an HMAC-signed notification the moment a task finishes — or use run-and-wait for fast tasks and get the result in the same call.
Batch & duplicate-safe
Submit up to 50 tasks in one call, and retry any request safely with an Idempotency-Key — a network retry never creates or bills a duplicate task.
Confidence, not just pass/fail
Every verified result carries a 0–100 confidence score, so you know how much to trust it — not just whether it technically ran.
Real-time usage
Check plan limits against real current consumption — agent count and task usage for the period — before you hit a wall, not after.
How completion is verified: Task completion is gated on verification evidence, not a model's own confidence — a task is only ever marked completed when its output has been checked against the original goal. Memory is strictly isolated per customer account; nothing is ever shared or reused across organizations.
Use cases
- Support triage agent: reads an incoming ticket via your webhook, classifies it, and drafts a response.
- Research agent: given a topic, gathers information via HTTP tools and returns a verified summary.
- Sales ops supervisor: delegates lead enrichment to one specialist and outreach drafting to another, then verifies both before handing off a ready-to-send package.
- Data pipeline agent: calls your internal APIs in sequence, retries transient failures on its own, and reports exactly what happened.
Plans
Pro
Multi-agent orchestration and your own custom tools.
$199/mo
- Up to 15 agents
- 5,000 tasks/mo
- Multi-agent delegation (supervisor + specialists)
- Custom HTTP tools
- Everything in Starter
Business
Volume workloads with unlimited agents.
$599/mo
- Unlimited agents
- 50,000 tasks/mo
- Priority rate limits
- Everything in Pro
FAQ
What happens when a task fails?
The agent diagnoses the failure, tries a genuinely different approach, and retries — up to a hard limit you set (default 3 attempts). It never loops forever, and it never marks a task completed without a separate verification pass confirming the result actually satisfies the goal.
Is memory shared between customers?
No. Every memory entry is scoped to your organization and the specific agent that created it. It is never queried across accounts, and you can disable or clear it per agent at any time.
Can one agent delegate to others?
Yes, on the Pro plan and above. A "supervisor" agent can delegate sub-tasks to "specialist" agents and verifies their combined results before returning a final answer.
What tools can an agent use?
Any HTTP API or webhook out of the box — including your own. Register a custom tool (a name, URL, method, and headers) and any agent can call it by name. Named connectors for Slack, GitHub, CRMs, and email are on the roadmap; today they route through the same generic HTTP mechanism.
Can an agent use a tool or write action I didn't grant it?
No. Tool access and write permissions are enforced when the task actually runs, not just suggested to the model — a plan that names a tool outside the agent's grant, or a non-GET call without the http:write permission, is refused before it ever executes.
What if my request times out and I retry it?
Send an Idempotency-Key header. Retrying the same request with the same key returns the original task instead of creating and re-running a duplicate — no accidental double charges against your task quota.
Do I have to poll for results?
No — pass a webhookUrl and we'll POST the final, HMAC-signed task state to it once it completes or fails. Polling GET /tasks/{id} still works if you prefer it.
Is this specific to Amber's own systems?
No — this is a general-purpose agent execution API for any developer's tools, data, and workflows.
Get your first agent running in minutes.
Subscribe, create an API key, and create your first agent — no sales call.