A photographer builds a DSL for multi-agent workflows

A thread appeared on Hacker News on May 31, 2026 with barely 3 points and no comments: a photographer publishes AgentFlow, a custom domain-specific language (DSL) for defining and orchestrating multi-agent workflows. No corporate backing, no engineering team behind it. Just someone who uses agents in their daily work and decided that existing tools didn't fit how they think.

That a project like this comes from outside the usual circle of professional software developers isn't the story itself. The story is that it's already normal.

What AgentFlow is and what it proposes

According to the repository, AgentFlow is a domain-specific language designed to describe pipelines where multiple agents collaborate, delegate tasks, and share context. The core idea is to separate the definition of the workflow (which agent does what, in what order, under what conditions) from the implementation of each individual agent.

It's a declarative approach: instead of writing imperative code that calls an agent, waits, processes the response, and calls the next one, the developer describes the topology of the workflow in the DSL and a runtime executes it. Conceptually this isn't new, with precedents in microservices orchestration and tools like LangGraph, but the value proposition here is simplicity and a focus on readability for non-engineers.

The repository includes examples of workflows with conditional branches, loops, and artifact passing between nodes. There's no explicit mention of integrations with MCP servers or Claude Code in the visible documentation at time of publication, though the architecture is compatible with any agent that exposes a standard calling interface.

Why it matters that a photographer built it

This isn't a comment on the author's credentials. It's an observation about the ecosystem: abstractions for working with agents have dropped low enough in level that someone without formal training in software engineering can build their own orchestration language and publish it.

Two years ago, coordinating multiple agents required a solid understanding of state management, handling network errors, serializing context, and designing communication interfaces between processes. Today, with tools like Claude Code, which already supports invocable sub-agents, lifecycle hooks, and reusable skills, along with MCP as a standardized integration layer, the plumbing work has dropped dramatically. What remains is the flow logic, and that's where projects like AgentFlow find their niche.

The author's profile also points to a concrete use case: workflows for processing, cataloging, or automatically describing images, likely with agents specialized in different parts of the pipeline (metadata extraction, description generation, classification by style, etc.). Not stated explicitly, but it's the most obvious reading.

Clear limitations and signals to watch

With 3 points on Hacker News and zero comments, AgentFlow hasn't generated public traction yet. The repository is recent and documentation is sparse. There are no visible tests, no packaged version, and no detailed installation instructions. In its current state, it's a personal project released publicly.

What deserves attention isn't whether AgentFlow succeeds as a tool, it may not, but the pattern it represents: non-technical domain users building their own abstraction layer on top of agent infrastructure. When that starts happening repeatedly in a community, it's usually a signal that the official abstraction layer still has gaps.

If the author continues developing the project and adds integration with MCP servers or support for Claude Code hooks, it could become something interesting for small teams that want to define agent workflows without writing orchestration code by hand.

---

From our perspective, the context matters more than the project itself: when domain users start building their own orchestration tools, the ecosystem has truly matured. It will be interesting to see if Anthropic picks up on this signal in future versions of Claude Code.