AWF – AI Workflow Framework

AWF, an open-source AI workflow framework, launched a three-part toolset: the Go-powered AWF CLI for YAML-defined, provider-agnostic workflows across OpenAI, Claude, Gemini and Ollama; ZPM, a Zig and Prolog reasoning layer exposing 22 MCP tools for facts, rules, queries, proofs and truth maintenance; and ztick, a Zig push-driven scheduler that fires jobs at requested times over TCP, webhooks, AMQP, Redis or AWF workflows. The framework claims deterministic replay of CLI workflows on identical inputs and provable ZPM deductions through Prolog inference chains, with per-step, per-workflow and per-run token tracking. AWF ships as static Go and Zig binaries with zero runtime dependencies, a native TCP protocol with TLS 1.3 and REST/OpenAPI 3.1.1, and positions itself as OSS end-to-end with no vendor lock-in.

AWF CLI orchestrates your AI workflows. ZPM gives your agents a memory they can reason over. ztick fires them at the precise moment you decide. Use one, use all three — they're designed to work together. 01 — The Orchestrator CLI Go-powered, provider-agnostic, YAML-native. Define workflows once, run them from your terminal, wire them into CI. The foundation everyone builds on. Zig + Prolog. Give your LLMs deterministic logical reasoning. 22 MCP tools for facts, rules, queries, proofs, and truth maintenance — usable as MCP server or direct CLI. Zig-powered, push-driven scheduler. Cron in reverse: your apps send SET over TCP at the moment they decide, ztick fires at the requested time — into shells, webhooks, AMQP, Redis, or AWF workflows. Native TCP protocol with TLS 1.3 + REST/OpenAPI 3.1.1 Vendor-agnostic at the orchestration layer — OpenAI, Claude, Gemini, Ollama — and protocol-open at the reasoning layer via MCP. Zero lock-in, OSS end-to-end. 02 Deterministic AI CLI workflows replay identically on the same inputs; ZPM deductions are provable through Prolog inference chains. AI becomes as predictable as the rest of your stack. 03 Production-Grade Go and Zig, static binaries, zero runtime dependencies. Structured logs, audit trails, and proof explanations across both products. Built for enterprise, not demos. Chapter 03 — Why It Matters What you actually get Concrete problems the ecosystem solves, in plain terms. 01 Ship in days, not months Workflows in YAML, one static binary, zero SDK sprawl. Install, write, run — today. 02 Your AI bill stops being opaque Every token tracked per step, per workflow, per run. Audit spend line by line — catch runaway agents before they run away. 03 Swap providers without a rewrite OpenAI raising prices? Switch to Claude. Need privacy? Drop in Ollama. Edit one line of YAML — your workflows don't care. 04 Explain every AI decision to audit CLI workflows replay identically. ZPM proves each conclusion through explicit inference chains. "Why did the AI say X?" finally has an answer. 05 Stop your agents from making stuff up Give LLMs a knowledge base that reasons. Facts persist, rules apply, contradictions surface. "The AI hallucinated" becomes "the assumption was wrong". 06 Run in prod without a DevOps team Two static binaries, zero runtime dependencies. Docker if you want, bare metal if you prefer. Any CI, any server, any laptop. …or, let's be honest The reasons you'll actually end up installing it tonight. Your engineers were tired yesterday and prompted worse. You read a wild thread on Reddit last Friday at 1 a.m. Your last AI demo segfaulted in front of the CEO. Accounting cried at the last vendor invoice. Your prompt is now 4,000 lines and nobody owns it. “ChatGPT said so” is not an architecture diagram. The all-hands deck has the word agentic on slide 3. Chapter 04 — The CLI AWF CLI The foundation of the ecosystem. A lightweight, powerful tool to build and run AI workflows from your terminal. 01 Lightning Fast Written in Go for blazing performance. Every millisecond counts in production. 02 Provider Agnostic Switch between OpenAI, Anthropic, or local LLMs without changing a line of code. 03 YAML Native Define workflows in simple, version-controllable files. Human readable, machine precise. 04 Production Ready Built-in validation and error handling for mission-critical deployments. Define your entire workflow in a single YAML file. Human-readable, version-controllable, and production-ready. Built for pros When the demo ends and production begins Loops For-each & while Iterate over arrays, file lists, or API results with full workflow context inside each iteration. type: foreach Conversations Multi-turn agents Native session resume on Claude, Codex, Gemini, OpenCode. Token tracking and context injection across turns. type: conversation Resume Pick up where it crashed Every run has an ID. Interrupt it, fix it, resume from the failed step — no need to start over. awf resume <id Sub-workflows Compose like functions Call one workflow from another with typed input/output mapping. Build libraries of reusable pipelines. type: call workflow Parallel Fan out without fan-pain Run multiple steps concurrently with configurable strategies — all, any, race. Collect results in one place. type: parallel Retry Backoff, baked in Exponential, linear, or constant backoff per step. Flaky APIs and rate limits stop waking you up at 3 a.m. retry: {strategy: exponential} Diagrams Export the state machine Generate Graphviz DOT for any workflow. Render it, ship it to architecture docs, impress the auditor. awf diagram <workflow Audit trail JSONL, secrets masked Paired start/end entries per step, atomic writes, secret redaction, replayable history. Compliance happy. awf history See it in action Execute your workflows with a single command. AWF provides detailed logging, step tracking, and clear completion status with workflow IDs for reproducibility. Terminal bash $ awf run code-review --input file=service.go --output=streaming --verbose=true registered desktop notification backend registered webhook notification backend Running workflow: code-review starting workflow workflow=code-review id=3eb274fa-186e-4070-a6a0-c26eea61ad28 executing step step=read executing step step=analyze executing agent step step=analyze provider=claude executing step step=report OUT Analysis complete recorded execution history workflow id=3eb274fa-186e-4070-a6a0-c26eea61ad28 status=completed workflow completed step=done status=completed ✓ read: completed successfully Workflow completed successfully in 36.6s Workflow ID: 3eb274fa-186e-4070-a6a0-c26eea61ad28 --- Execution Details --- Status: completed Steps executed: report: completed 1ms read: completed 1ms analyze: completed 36.585s AWF provides detailed execution logs, validation feedback, and clear success/failure status for every workflow run. Get Started with AWF CLI The core engine of the AWF ecosystem. Install it in seconds and start building. If it's in your PATH, it's in your workflow. The CLI orchestrates any CLI utility with zero configuration. Chapter 05 — Extend It Plugins & Workflow Packs The CLI is a platform, not just a tool. Write your own step types or share a whole workflow as a reusable pack. Plugins Write your own step types Plugins add new primitives to the workflow DSL — inject real-time data, call internal APIs, run custom tools. Distribute them as binaries, load them from anywhere. Give your LLM a memory that reasons. ZPM exposes the same 22 tools over the MCP protocol and as direct CLI subcommands — use it from Claude Code, Cursor, or straight from your shell. Get Started with ZPM The reasoning engine of the AWF ecosystem. One binary, zero runtime dependencies. The triggering layer of the ecosystem. A Zig-powered, push-driven scheduler — cron in reverse. Apps schedule jobs at the moment they decide, ztick fires at the requested time. 01 Push-Driven Your app sends SET over TCP at the moment it decides. ztick fires at the requested time. No polling, no fixed intervals, no drift. 02 Multi-Runner Dispatch Shell, direct execve, HTTP webhook, AWF workflow, AMQP publisher, Redis command. One scheduler, every backend you already run. 03 TCP + REST + TLS Native TCP protocol with TLS 1.3, plus REST/JSON HTTP API with embedded OpenAPI 3.1.1. Token AUTH, namespace-scoped authorization. 04 Durable & Observable Append-only logfile with binary framing and background compression. OpenTelemetry traces and metrics over OTLP/HTTP. Zero runtime dependencies. A tight TCP protocol or a REST endpoint — pick your interface. ztick chains directly into AWF workflows, so you can trigger orchestrations on demand: no cron, no drift, no surprises. Get Started with ztick The triggering engine of the AWF ecosystem. One Zig binary, zero runtime dependencies.