OpenAI Codex Lead: Why Juggling 10 AI Agents Isn't the Future

The Attention Bottleneck

In an August 24, 2026 interview with Matthew Berman, OpenAI Codex lead Tibo describes his current workflow: he often launches 10–15 agents in parallel, each taking 30–45 minutes to return results. One agent waits for permissions, another delivers a diff, a third hits a dead end and asks whether to change approach. The developer no longer writes code line by line but constantly context-switches between tasks.

Tibo: "Managing your attention." Tibo: "先把人的注意力管理好。"

Tibo emphasizes that model capabilities are growing — tasks get longer — but human attention hasn't scaled. The problem shifts from "can we run more agents?" to "who manages them?" He explicitly rejects the idea that juggling many agents is an advanced workflow: "I was multitasking like 10 agents. I don't want to really go back to that."

Current Tooling Gaps: Skill, Memory, Subagent

Today's coding agents already include Skill (saved work methods), Memory (project conventions), and Subagent (parallel execution). Power users maintain rule files, compress context, and split sub-tasks. Yet Tibo is unsatisfied:

Skill files drift out of date as directories change.

Memory doesn't reliably retain all facts.

Enabling Subagents forces users to orchestrate division of labor themselves.

Repositories accumulate AGENTS.md, rule directories, prompt templates, hook scripts, and MCP configs. These boost success rates but transfer "AI environment maintenance" burden to developers. Examples: a Skill references an old path after a restructure; Memory cites a deprecated module; two Subagents edit the same file, leaving conflict resolution to the human.

Tibo: "Deeply understands you." Tibo: "真正理解你。"

He wants the product to internalize these details: understand the user's goals, daily work, and team context — not start from a blank chat each time.

Vision: Unified Entry Point with System Orchestration

Discussing Codex and ChatGPT convergence, Tibo notes early users wanted separate interfaces (terminal/diff/tests vs. docs/web). His judgment: models are flattening that boundary.

Tibo: "The same harness." Translation: "底层会使用同一套Agent执行系统。"

In a team setting, a single entry point doesn't mean one model. The backend can dispatch different agents: one searches the repo, one reproduces a bug, one reads logs, one runs tests. The user sees only task progress and final results. This orchestration layer must retain three things:

Task state

Context provenance

Completion criteria

Risk-Based Human-in-the-Loop

A multi-agent system's easiest mistake is asking the human for every trivial decision. Ten agents each prompting "continue?" would stall the developer all day. Tibo proposes tiered risk handling:

Auto-approve: read repo, search symbols, run local tests.

Logged & scoped: modify dependencies, write to shared branches, access test databases.

Human confirmation required: production deployments, secrets access, payments, data deletion.

Tibo: "A high-risk action." Translation: "只把高风险动作交给人确认。"

Notifications follow the same logic: a failing test that the agent can self-repair doesn't need an immediate alert; a requirement drift or irreversible operation does. Useful notifications tell the developer: what diverged from the goal, what's next, and whether the step is reversible.

Community Experiments: Building the Orchestration Layer

Developers are already prototyping this layer. One shared a solo system where Codex and Claude Code handle execution, while a scheduler above them stores business context, assigns tasks, monitors progress, and notifies only when a PR meets merge criteria. This reveals real constraints: task boundaries still need human definition; each agent in its own worktree consumes memory rapidly; over-permissioned schedulers become new security surfaces.

Another discussion treats rules, Skills, hooks, and Subagents as engineering components to stabilize model behavior and reduce context drift.

Both community efforts and Tibo's vision converge: multi-agent isn't just opening more chat windows. It requires a task system, a permission system, and an acceptance system. Without them, the human remains a message sorter.

Conclusion

Running 15 agents looks like stuffing a dev team into a laptop. As results stream back, the human drowns in window-switching, context-rebuilding, permission-granting, and diff-reviewing. Tibo's future isn't "one programmer commanding an agent army" — it's the opposite: the system understands the goal, arranges background work, decides when to proceed autonomously and when to escalate. The developer keeps ownership of goals, boundaries, and final judgment. What disappears is the busywork of context switching, repetitive briefings, and meaningless approvals. When agents start managing agents, AI programming graduates from "spin up more assistants" to a genuinely new way of working.