Working Notes on Agent Systems/Brad Zhang

@teach_fireworks / X longform

Codex’s weekly active developers exceed 4 million! Codex is definitely OpenAI's most po...

Codex’s weekly active developers exceed 4 million! Codex is definitely OpenAI's most popular chicken. It has reversed the decline in one fell swoop and turned from defense to of...

May 22, 2026 · 3 min read

Codex’s weekly active developers exceed 4 million!

Codex is definitely OpenAI's most popular chicken.

It has reversed the decline in one fell swoop and turned from defense to offense.

It has eaten up a very large share of ClaudeCode in more than a month.

Don't think that the impact is not big, it will slowly change the usage habits and mentality of C-end users, thereby slowly affecting the to B business.

In addition, the company has IPOed this year in the past two years, and the competition has reached a fierce stage.

The intelligent agent ushered in a long-term running state.

OpenAI launched the Codex target mode on Thursday, covering the application end, integrated development environment and command line interface.

The interaction logic has completely changed: there is no need to write instructions repeatedly, just issue the final mission goal directly.

Earlier this week, Codex was connected to the ChatGPT mobile terminal, and the agent can continue to work even after turning off the computer.

Official simultaneous disclosure: Codex has more than 4 million weekly active developers, making it OpenAI’s fastest-growing enterprise-level product.

Anthropic simultaneously increased its layout, and Andrej Karpathy led the team to rely on Claude to iterate its own model training, and recursive AI research and development was upgraded to an independent department.

Google launched Gemini Spark at its developer conference on Tuesday, equipped with a dedicated cloud virtual machine and supporting continuous work in the offline background.

The three leading companies are in sync: intelligent agents are evolving into permanent work units.

The user only needs to deliver the task target, and the AI ​​will complete it in a closed loop autonomously.

The thinking logic of the industry has been iterated: in the past, we struggled with where to use code assistants, but now we need to clarify which tasks in the enterprise can be directly completed by AI, and which ones still require manual follow-up.

Agent still has a long development track.

SaaS-Bench: Complete cross-application full-process practical tasks in real SaaS systems.

Even for top agents, the end-to-end task completion rate is less than 4%.

The reasons for failure are all common problems in the industry: broken task planning, loss of cross-application status, and inability to self-repair after errors.

At the same time, METR released a cutting-edge risk report and carried out actual tests on self-developed agents from Anthropic, Google, Meta, and OpenAI.

Compared with the ranking, the test conclusion is more critical: the core bottleneck of the current implementation of intelligent agents is no longer the upper limit of model intelligence, but depends on whether humans can effectively supervise it.

In scenarios where the results are easy to verify, the performance of the agent is stable; once the verification process is complicated, the operation can easily crash.

Visual summary

Article argument map

Generated from the post's content graph

FORMATTOPICCAPABILITYMARKETcoverscoverscoverscoverssignalssignalssignalssignalsFORMATlongform noteTOPICagentsTOPICtechnical distributionTOPICretrievalTOPICtoolingCAPABILITYagent workflowCAPABILITYdeveloper toolingCAPABILITYproduct surfaceCAPABILITYtechnical distributionMARKETopen-source builders
Mermaid outline
flowchart LR
  format-long_post["longform note"]
  topic-agents["agents"]
  topic-technical-distribution["technical distribution"]
  topic-retrieval["retrieval"]
  topic-tooling["tooling"]
  capability-agent-workflow["agent workflow"]
  capability-developer-tooling["developer tooling"]
  capability-product-surface["product surface"]
  capability-technical-distribution["technical distribution"]
  market-open-source-builders["open-source builders"]
  format-long_post -->|covers| topic-agents
  format-long_post -->|covers| topic-technical-distribution
  format-long_post -->|covers| topic-retrieval
  format-long_post -->|covers| topic-tooling
  format-long_post -->|signals| capability-agent-workflow
  format-long_post -->|signals| capability-developer-tooling
  format-long_post -->|signals| capability-product-surface
  format-long_post -->|signals| capability-technical-distribution

Visual structure

Essay structure map

Built from summary and key paragraph positions

Codex’s weekly active developers exceed 4 million! Codex is definitely OpenAI's most...THESISCodex’s weekly activedevelopers exceed 4million! Codex isdefinitely OpenAI'sSIGNALCodex’s weekly activedevelopers exceed 4million!OPERATORAnthropicsimultaneouslyincreased its layout,and Andrej KarpathyIMPLICATIONIn scenarios where theresults are easy toverify, theperformance of the
Mermaid outline
flowchart LR
  thesis["Codex’s weekly active developers exceed 4 million! Codex is definitely OpenAI's most popular chicken. It ha..."]
  signal["Codex’s weekly active developers exceed 4 million!"]
  operator["Anthropic simultaneously increased its layout, and Andrej Karpathy led the team to rely on Claude to iterat..."]
  implication["In scenarios where the results are easy to verify, the performance of the agent is stable; once the verific..."]
  thesis -->|frames| signal
  signal -->|develops| operator
  operator -->|lands in| implication

Source: View the original post