Building with AI agents is amazing until they run out of context, tokens, and break alignment on complex tasks.
We struggled with this a lot while constructing our own agent coding factory. After months of trial and error, we finally built the right internal tools and fine-tuning processes to keep our agents efficient for long-duration tasks. We’d love to share our story, show you the tools we built to solve our own headaches, and help your team bypass the struggles we faced!

We built a NanoClaw agent factory where autonomous agents triage, review, spin up test VMs, and merge every pull request—even prompt injections. I’ll explain its three safety layers: OS-level sandboxing, a zero-touch credential vault, and identity-bound human approval gates. Finally, I’ll demonstrate how our supervisor, reviewer, and test agents coordinate in Slack, before letting the audience interact live with my personal AI assistant.

When my team started using coding agents we felt the magic… until we scaled their use. Then the cracks started to show:
– Which environment did the feature land in?
– Monotonous tracking of which new features are up and running to develop against
– Architectural decisions didn’t stick for the next session.
– ExpiringMissing context made shipping a working feature take longer.

It all came back to the same thing: our agents were shipping code faster than the rest of the software lifecycle could keep up. Those pains sent us on a journey to build Fly.

Fly is an agentic platform that stores your packages, manages your releases, and tracks them in runtime, all from your coding agent. Fly also auto-tracks the decisions you and your agent make during a coding session, binds them to the release you ship, and surfaces them to the next teammate’s agents automatically.

During this session, I’ll share our experience with coding agents and how we solved some of the biggest pains a team faces when working with them to collaboratively release and ship real software.

Building with AI agents is amazing until they run out of context, tokens, and break alignment on complex tasks.
We struggled with this a lot while constructing our own agent coding factory. After months of trial and error, we finally built the right internal tools and fine-tuning processes to keep our agents efficient for long-duration tasks. We’d love to share our story, show you the tools we built to solve our own headaches, and help your team bypass the struggles we faced!