PowerPoint Subagent¶
Concrete reification of subagent-architecture at listen-labs. Florian Juengermann rebuilt Listen's slide-deck generator from scratch on top of the claude-code SDK, running inside an E2B sandbox, because the original custom pipeline couldn't hit the iteration-quality bar.
How it works¶
- Main research agent decides "we should produce a slide deck summarizing this analysis."
- It spawns the PowerPoint subagent with: (a) the content spec, (b) the customer's brand template, (c) a skill describing how to modify a
.pptxfile in Python. - The subagent — which is literally Claude Code's agent loop running inside E2B — writes Python that loads the template, injects content, and iterates over multiple passes (checking its own output, re-rendering, fixing layout).
- The rendered deck returns to the user.
Florian's account (source):
"We have this cloud code agent SDK that runs and writes Python code and can then modify the PowerPoint file and kind of do a lot of iterations in that… it's a sub-agent to our main agent."
Why a full rewrite¶
The original custom pipeline existed. They threw it out for Claude Code SDK because: - The iteration loop (write code → run → inspect output → fix) was already solved by Claude Code - Maintaining their own harness for one niche task wasn't justified once the SDK was available - Brand templates have enough fiddly one-off edge cases that a general code-writing loop beats a specialized template engine
Related¶
- skill-distillation — Listen passes a "PowerPoint skill" to the subagent so it doesn't reinvent the SDK
- subagent-architecture — generic pattern
- claude-code — the underlying harness being embedded
- isolates-vs-containers — E2B is containers-route for this sandboxing