OpenAI's Codex deployment marks a turning point. 10,000 agents, zero human PRs in 72 hours. Here's what this means for the $4T software industry.
The agents are writing, testing, and merging code across OpenAI's entire production codebase. Senior engineers are now agent supervisors.
RAG-based memory works for retrieval but agents lose taste. Here is a structured scratchpad approach that solved my 6-hour monorepo refactor agent.
Both agents were given the same 50 tasks on our production monorepo. Codex was faster, but Claude was more careful.
The agent economy just crossed $200B in market cap. But the real question is about the fundamental redefinition of employment.
Apache 2.0 license. Runs on a single A100 with 4-bit quantization. Beats GPT-4o on MMLU, CodeContests, and HumanEval.
The first fines are expected by Q4 2026. This is a practical technical guide to the new requirements.
One startups Pinecone bill went from $200/mo to $18,000/mo. Alternatives: Qdrant self-hosted, pgvector, and the just use BM25 camp.
We spent 2 weeks testing all four on real projects. Each has different strengths.