What I'm watching in AI×GTM

Amjad Masad and Me at SaaStr AI 2026: The Agents We Actually Built, and What Replit’s Founder Thinks Comes Next

Especially the overall idea of reporting to agents

• Context windows expanding from 16K to 1M+ tokens enables perpetually-running agents that never need rebooting, fundamentally changing agent architecture from ephemeral to persistent
• Mono repo architecture (10 apps in one codebase) beats separate apps for AI agents because global context compounds - agent remembers how it built previous apps when building new ones
• Self-improving agents are production-ready: Replit's internal agent autonomously reads traces nightly, generates PRs with prompt improvements, ships A/B tests, and loops back without human intervention

GLM 5.2: why I’m replacing Opus in Claude Code with this new model

• Open-weight models like GLM 5.2 can replace premium models (Claude Opus) at 1-2% of the cost ($3.36 for 6M tokens vs ~$300+ for Opus)
• GLM 5.2 successfully handled production tasks: codebase architecture audit, UI redesign matching existing design system, and 45-minute autonomous bug hunting from real logs
• Open-weight models provide vendor independence and cost predictability for AI coding workflows, challenging the assumption that frontier models are necessary for production coding tasks

I added a clause to Andrej Karpathy's 4 CLAUDE.MD clauses for Claude Code. It has been a game changer for me.

• Karpathy's original 4 Claude rules (ask don't assume, simplest solution first, don't touch unrelated code, flag uncertainty) were designed to prevent AI hallucination and over-engineering but had unintended consequence of silencing AI's reasoning capabilities
• Adding a 5th clause explicitly inviting Claude to suggest better approaches transformed it from a 'code producer' back into a 'pair programmer' that contributes strategic thinking during code reviews
• The evolution shows a maturation in AI coding workflows: early adopters started with strict constraints to prevent bad AI behavior, now refining to unlock collaborative reasoning while maintaining guardrails

How Claude Mythos found a 15-year-old bug in Mozilla Firefox | Brian Grinstead

• Agentic bug-finding requires sophisticated infrastructure—LLM judge for file ranking, verifier subagent to catch false positives, goal-loop pattern for retries—not just pointing AI at code
• Teams with existing fuzzing, CI, and dev tooling infrastructure have massive advantage in AI adoption; the harness matters as much as the model
• The 'score, verify, fix' loop pattern is generalizable beyond engineering to design quality, conversion optimization, and tech debt—non-engineers can reuse the framework