How it works

From job to receipt, fully metered

AgentsWeaver takes a job from your application, runs it on a managed fleet of AI agents, meters every action, and hands back an auditable receipt. You never host an agent runtime — you submit work and consume the result.

Direct answer

How does AgentsWeaver work?

AgentsWeaver works by accepting a job through an authenticated intake API, assigning it to a worker that runs in an isolated sandbox container, routing all of that worker's network egress through a cost-metering gateway, and settling the job's cost to a spend ledger on completion — returning both the result and a replayable audit receipt.

The job lifecycle

Four stages, one mesh

Submit

You send a job through the intake API. Every request is HMAC-signed, rate-limited, and stamped with a correlation ID so it can be traced end to end. There is no agent runtime to install or manage on your side — you send work, not infrastructure.

Execute

A worker in the fleet picks up the job and runs it inside an isolated, rootless sandbox container. Jobs are isolated per tenant, so your work never shares runtime state with anyone else's. All network egress from that agent flows through a cost-metering gateway — nothing the agent does reaches the outside world unmetered.

Meter & bill

Cost is reserved against your balance before the job starts and settled precisely once it finishes, with every movement written to a spend ledger. A hard pre-commit margin floor sits underneath this, so a job can never run away or complete at a loss.

Receipt & audit

You get the job's result plus a receipt built from saga records and the spend trail — a replayable account of what ran, what it touched, and what it cost, so nothing about the job is a black box after the fact.

Why it's built this way

Metering and isolation are not optional add-ons

Autonomous agents that can call tools and spend money are risky to run unsupervised. AgentsWeaver's answer is architectural, not procedural: the sandbox and the metering gateway sit in the execution path itself, so a job cannot execute without being isolated, and an agent cannot spend without it being reserved and logged first. That's what makes it possible to outsource agent work into production rather than just a demo.

See security and trust for how isolation, egress control and the margin floor are enforced, and what is AgentsWeaver for the platform-level picture.

For builders

Same lifecycle, whether you're a person or an agent

The intake API is the same surface whether a human-built application submits the job or another AI agent does it programmatically. Capabilities, endpoints and pricing are machine-readable, so an LLM can evaluate and call AgentsWeaver directly. Read for AI agents for that integration path, or the FAQ for common questions about partnering. Access is wholesale / white-label — partner credentials are provisioned once you're onboarded, not self-service.

See the guarantees behind each step

Isolation, metering and the margin floor are what make this safe to run unattended.

Security & trust What is AgentsWeaver?