Skip to content

Azhi Flow

Governed, durable agent workflows that say plainly what they can and cannot guarantee. Every run shows what the agent saw, what it was allowed to do, what it did, and what it cost.

Most agent tools start from an agent and hand it tools. Azhi starts from a workflow, and an agent is only one kind of node in it. Fetching data, parsing, filtering and posting run as ordinary code. A model is called only where judgment is needed. The whole thing runs durably on Temporal, behind a tool gateway that records every write, and it tells you before the run what is enforced and what is not.

Durable runs

Crash, restart or timeout and the run resumes. Retries by error class, cancellation, schedules in IANA timezones, and waits for offline workers. Built on Temporal.

A plan before anything executes

azhi plan marks every requirement native, bridged, unsupported or unverified, and every policy enforced, harness-enforced or unobservable. A run with blockers is refused.

Exactly-once writes

The tool gateway keeps an action ledger. Kill the process mid-Slack-post and exactly one message exists, and azhi inspect shows why.

Prompt-injection guard

An ungated write after an agent that read untrusted data fails to compile. It is caught at authoring time, not in production.

Honest cost accounting

Usage per turn, cost from declared pricing or “unavailable”. Never guessed. Per-step budgets and daily/monthly spend limits.

Bring your own harness

Built-in model agent, OpenCode, Claude Agent SDK. Each declares what it can enforce and the plan checks it. Anthropic, OpenAI and GitHub Copilot models.

PlatformSupported as
LinuxServer and workers. The reference platform.
Windows 10/11Everything inside WSL2, or the experimental no-Docker local mode natively.
macOSServer in Docker plus a development worker. Tested and working; workers not certified.