durable agent runtime

Zeta

Author agents as files. Run them as durable events. Inspect every model call, tool result, queue claim, and prompt boundary.

Deploy an agent Inspect the runtime

01 authored specs become explicit event routes

02 queues, leases, and locks are durable tables

03 prompt provenance is replayable by construction

agents/ valid

release-manager.md

support-triage.md

skills/

code-review.md

release-notes.md

events/

github.pr.opened.json

release.ready.json

tools/

read.py

grep.py

github.pr.opened

accepted by release-manager, projected into one queue item

runtime.attempt.running

lease token fenced, heartbeat active, locks acquired

release.ready

returned event is schema-checked before it composes

zeta run --forever live

$ zeta run --project-root .

accepted github.pr.opened

queued qi_evt_421_release-manager

claimed worker=local-runtime

lock repo:release acquired

model local-model streamed 822 tokens

tool grep returned 12 grounded lines

completed att_qi_evt_421_1

emitted release.ready

queueavailable

attemptcompleted

tracereplayable

schemavalid

runtime

Agents that behave like software.

Zeta treats every transition as a durable boundary: incoming events, queue lifecycle, model calls, tool results, prompt components, and returned events all land in an inspectable log.

01 / authored

An agent is a contract.

Frontmatter declares what an agent accepts, what it may return, which tools it can reach, and whether it can resume state.

02 / durable

The queue is visible.

Claim tokens, attempt rows, locks, heartbeats, and terminal writes are explicit runtime facts instead of hidden control flow.

03 / bounded

Capabilities stay fenced.

Tools are registered as contracts. Staged effects remain inspectable before they touch the outside world.

trace graph

Prompt provenance without folklore.

Every answer can be traced back to its prompt components, tool calls, model output, and durable event chain.

Replay the prompt.

Reconstruct the exact model request from stored components, selected tools, token budgets, and current timeline.

Diff the answer.

Compare two runs by prompt object, tool result, model output, and returned event instead of guessing from logs.

Trust the boundary.

Durable event schemas make runtime behavior inspectable before agents compose into larger systems.

authoring

Markdown for policy. Events for motion.

Zeta keeps the agent surface small: a Markdown file, explicit event schemas, reusable skills, and a runtime that records what happened.

---
name: Release Manager
description: Prepare release summaries.
accepts:
  - github.pr.opened
returns:
  - release.ready
tools:
  - read
  - grep
skills:
  - code-review
schedules:
  - cron: "0 9 * * 1"
    timezone: Europe/Paris
resumable: true
---
Review {{ event.payload.title }} and emit a release summary.

eventschema

queuelease

prompttrace

resultcontract

No
Bad
Runs

start

Ship the agent loop you can audit.

$ mkdir -p agents/events agents/skills

$ edit agents/release-manager.md

$ zeta events publish github.pr.opened --json payload.json

$ zeta run --forever