@mfpiccolo: https://x.com/mfpiccolo/status/2064358779940995141
Summary
The article introduces iii, an open-source platform for building production-ready agentic backends using composable workers that replace traditional service integration, offering a React-like paradigm for backend engineering.
View Cached Full Text
Cached at: 06/10/26, 07:48 AM
How to make your Agentic Backend Architecture Production Ready?
Building an agentic backend on iii means composing a handful of workers on the engine. Each one is production-ready. Each one is replaceable. The composition is your backend.
If you see below in the image, harness sequences the agent turn loop. context-manager turns raw history into model-ready context. session-manager is a durable, reactive, branching store of typed conversation entries. llm-router is one front door over every LLM provider.
Each one is standalone. Install just llm-router and you have provider-agnostic completions for any app. Install just session-manager and you have a reactive conversation store. The harness only sequences the other three plus tool(i.e. function calls) dispatch. Anything that grows real logic, approval gating, spend budgets, multi-agent handoff, becomes its own sibling worker rather than complicating the harness.
From outside this looks like four layers. Inside there is no integration code between them. The harness calls the context-manager with iii.trigger(“context::sync”). The context-manager returns. The harness calls the llm-router. The llm-router streams over a channel. No orchestration framework to configure. No contract between layers to maintain. No service mesh to operate.
These four workers run on the same engine as queue worker, state worker, pubsub worker, observability worker, HTTP worker, sandbox worker, schedule/cron worker. Every worker inherits the rest, with zero integration cost. iii worker add <worker-name>, get the whole system surface. Adding queue support to your llm-router is calling iii.trigger(“queue::enqueue”). Adding scheduled retries is calling iii.trigger(“cron::schedule”). Adding observability is automatic. These are not separate integrations. They are functions on workers already running.
This is the @iiidevs pattern. We noticed complexity in backend engineering eventually demands abstraction, the same way @reactjs noticed it in UIs. React reduced UIs to Components. iii reduces backends to Worker, Trigger, Function.
The paradigm of “add another service and integrate it into all others” needs to end. Any capability you would have added as a service can be a worker on iii. The functionality is exposed as functions and triggers. What you get is a system that is composable, extensible, discoverable, and observable. One that gets more capable every time anyone adds a worker.
Join us in building the Agentic Backend that the modern world needs: discord.gg/iiidev
iii is open source. Get started at iii.dev/docs. The harness workers are at github.com/iii-hq/workers and the engine is at github.com/iii-hq/iii.
Similar Articles
@mfpiccolo: https://x.com/mfpiccolo/status/2060069083878408689
The article argues that current agent harness frameworks like LangChain and CrewAI bundle independent concerns into a monolithic block, leading to inflexibility. It introduces the iii engine, where each responsibility is a separate, swappable worker connected via a shared bus and a single trigger primitive, allowing developers to compose their own harness by swapping workers rather than forking a framework.
@ghumare64: This is one of the cleanest architectural piece you might ready today. Separate workers when composed together they for…
An architectural approach for building agentic backends using separate workers that compose together without integration code, running on a shared engine that provides queue, state, pubsub, observability, HTTP, sandbox, and cron capabilities.
@_overment: https://x.com/_overment/status/2076440928726708612
This thread explores rethinking MCP (Model Context Protocol) beyond simple service integrations, highlighting generalized capabilities like search and execute, as demonstrated by Kody, a personal assistant that runs on Cloudflare and can be used by any agent.
@svpino: After MCP, AG-UI is probably the agentic protocol that's growing the fastest. It's a lightweight event-streaming protoc…
Santiago (@svpino) highlights AG-UI as the fastest-growing agentic protocol after MCP, a lightweight event-streaming protocol for building user-facing AI agents with support for real-time updates, tool orchestration, shared mutable state, security, UI sync, and now threads for resumable conversations.
@irl_danB: everyone is building an agent or a tool you don't want an agent or a tool, you want a reactor I've been working on some…
A developer introduces a concept called a 'reactor' — an agent session DAG that maintains a memoized world-model using OpenProse markdown files and the openai-agents-sdk, drawing parallels to React and dataflow.