Skip to content

Blog

Sub-Agent Orchestration: Cost-Aware Delegation With Work Attribution

Claude Code recently shipped Agent Teams, an experimental feature where fully independent Claude Code instances communicate with each other via mailboxes and shared task lists. Teammates can challenge each other's findings, collaborate on problems, and coordinate through a team lead.

HtmlGraph's orchestration system solves a different problem. It's not about agents talking to each other; it's about dispatching the right agent for each task at the right cost, tracking what each one produces, and merging the results with quality gates.

YOLO Mode: Autonomous AI Development With Engineering Guardrails

The question every developer using AI coding tools eventually asks: "Can I just let it run?" The appeal of autonomous AI development is obvious: you describe what you want, walk away, and come back to a working feature. The fear is equally obvious: what if it makes a mess?

Claude Code's answer is auto mode, a server-side classifier that screens every tool call for security threats like data exfiltration, production deployments, and force-pushes to main. It's a good answer to the security question. But it's only available on Team, Enterprise, or API plans, requires admin enablement, and only works with the Anthropic API provider. And it doesn't address a different question entirely: is the AI doing quality development work?

HtmlGraph's YOLO mode answers the quality question.

Plan Mode: From Freeform Text to Structured, Critiqued, Human-Reviewed Plans

Claude Code has a built-in plan mode. You press Shift+Tab twice, Claude researches your codebase in read-only mode, and produces a plan as freeform text. You review it, approve it, and Claude executes. It works for simple tasks.

But when you're planning a multi-week initiative with 8 interdependent slices, design decisions that need human input, and architectural trade-offs that need scrutiny, freeform text isn't enough. The plan vanishes when the session ends. There's no structured approval flow. Nobody critiques the plan before you act on it. And there's no mechanism to wire the approved plan into executable work items.

HtmlGraph's CRISPI plan system fills these gaps.

Why I Switched from Python to Go for AI Development Tooling

HtmlGraph started in Python. The Claude Code SDK was Python-based, the hook system used Python scripts, and I was comfortable in the ecosystem. It worked — until the project outgrew it.

The breaking point people ask about is latency, and that was real: every Claude Code hook invocation spawned a fresh Python process with a ~500ms cold start. A session with 200 tool calls meant 100 seconds of pure hook overhead. The agent would visibly pause between actions.

But latency wasn't the only friction. Running tests was slow. Distribution meant explaining virtualenvs and Python versions. As the codebase grew, the development experience itself was becoming a drag, not just for the tool's users, but for me building it every day.

Introducing HtmlGraph: Local-First Observability for AI-Assisted Development

My background is data analysis, not software engineering. At Sunnova and SunStrong, I wrote Python scripts and built ETL pipelines because the work required it, not because I set out to be a developer. But I've never been able to leave tools alone. If something doesn't work the way I think it should, I want to change it.

AI coding tools made that possible in a way it wasn't before. With Claude Code and Codex, a data analyst can build real developer tooling, not just scripts. HtmlGraph is the result of that: a local-first observability and coordination platform for AI-assisted development, built by someone who needed it for his own workflow.

It stores everything as HTML files in your repo (work items, plans, session records), all human-readable, git-diffable, and version-controlled. No Docker, no external databases, no proprietary formats. Just a single Go binary and your git repo.

I had an employment gap due to work authorization issues that gave me several months to go deep on this. What started as curiosity turned into a real project.