AI at Panfactum

• No usage Q1 ‘25 -> 100% AI-driven development Q1 ‘26

• Better code, 10x faster, 100x cheaper

• Cut SWE headcount budget by 80% while increasing our roadmap goals

• Forward-looking hiring strategy: We will only hire engineers who can demonstrate mastery of their agentic AI toolchains to be 10x more productive than they were in 2025.

What Changed?

• Model task adherence (doubling ever 7 months)

• Agentic harnesses that actually work (Claude Code)

• Sandboxing to allow --dangerously-skip-permissions (full autonomy)

• Ancillary tools that 10x agent capabilities:

  • Exa / Parallel (web-search for coding agents)

  • Hooks (recursive self-improvement via static analysis)

  • Workmux (local agent multiplexing)

  • Structured planning

Where I Spend my Time

What’s Next (Tech)

• 24/7 uptime

• Multi-machine scale out

• Autonomous improvements

• Improved token / context management

• Custom harness

Current Limitations

• Token throughput

• Multiplayer collaboration

• Provider reliability

What’s Next (Macro)

• Rise of the expert generalist / decreasing demand for specialists

• Cost of software approaches 0 asymptotically

• Redefinition of roles and career paths in software engineering

Recommendations

• Download Claude Code and dedicate >25% of bandwidth to agentic AI upskilling

• Learn how to use lightweight editors and terminal multiplexing

• Don’t expect your workplace to provide you the resources you need for this dynamic environment

• Build your own workflows and harnesses (set yourself up to take them with you across jobs)

• Don’t overcomplicate things (e.g., don’t start with gastown)

References

• PRD Skill

• Claude Sandboxing with Bubblewrap

• Basic Linting Hook