Coding Agents & LLM Mechanics

• GLM-5.2 open weights — MIT-licensed frontier model with 1M context window and dual reasoning tiers (max vs high); direct Sonnet/Opus-tier competition at open-weights pricing.
• Sakana Fugu — full multi-agent orchestration system exposed as a single model API endpoint; abstracts orchestration complexity from the caller entirely.
• Context > model quality for enterprise coding — teams switching tools found well-scoped context beats raw benchmark scores; directly validates the "context window hygiene" investment.
• Token Gap Death Loop — observed failure mode where the model fills context faster than it produces useful output, causing spiral degradation; worth understanding for long-horizon agent runs.
• Using git to make LLM-assisted work accumulate — framing git commits as the durable memory layer so each session inherits prior decisions rather than re-deriving them.
• Tracking agent behavior changes over time — teams lack canonical tooling for behavioral regression across model/prompt updates; open problem worth watching.
• AI for reviewing huge diffs — argument that AI review shifts human reviewer role toward architectural judgment rather than line-by-line scanning; practical framing for agentic PR workflows.
• Cline and Roo Code ecosystem churn — community migrating away; Claude Code and Aider gaining ground as the consolidation picks.
• Aider vs Claude Code — comparison thread surfacing where each tool's context management and autonomy differ in practice.
• Claude Code vs Codex usage on $20 plan — Claude Code's heavier per-task token spend vs Codex's lighter model; relevant to evaluating cost routing for agentic loops.

Agentic SaaS

• Factory AI — full-stack agentic dev platform (plan → code → review → automate) targeting engineering teams; not a copilot bolt-on, positions as an autonomous dev team layer.
• MCP hallucinated-package catcher — 20% of ChatGPT-recommended packages don't exist; this MCP server intercepts install commands and validates against real registries before execution.
• MCP config security scanner — tooling to audit MCP server configs for malicious or overly-permissive tool definitions; relevant as MCP adoption scales.
• AI UI design tool with LLM-readable exports — demand signal for design tools that output structured formats consumable by downstream coding agents, not just human-readable PNGs.
• Agent 37 Cloud — per-customer agent provisioning (Hermes/OpenClaw flavors); productized multi-tenant agent infrastructure as a SaaS primitive.

Vibe Coding & Ecosystem Shifts

• Vibe coding then vs now — snapshot of how the zero-to-deployed loop compressed from weeks to hours via Bolt/Lovable/V0 + Cursor/Windsurf; useful framing for where the floor of competence now sits.
• Roo Code 3M installs → shutting down for Roomote — team pivoting from IDE extension to remote/cloud execution model; signals the industry belief that the execution environment, not the editor plugin, is the real moat.