Codex multi-model split: Astra signs off, Terra swarms, Sol executes
A codex launch with an adversary agent and Terra subagents splits security audit, parallel swarming and code fixes across three models using plain-text IPC, with Jev handling the no-thinker forks in milliseconds.
The Head of Codex at OpenAI just shared an official tip as GPT-6.1 Terra drops today The setup: split your workload across GPT-6.1 Sol, GPT-6.1 Terra, and GPT-6 Astra, and stop burning Sol context on tasks Terra can swarm at 340 tok/s Zero context pollution between sessions: independent subagents execute via plain-text IPC in separate worktrees without burning a single token on conversation history You just launch Codex like this: codex --model gpt-6.1-sol --agent adversary --subagents gpt-6.1-terra here's the split: → GPT-6 Astra architects invariant security: audits ~/.codex/agents/adversary.toml, enforces sandbox policies, and signs off on multi-worktree diffs → GPT-6.1 Sol executes code fixes: scopes architecture, patches syntax, and merges pull requests whenever a worker reports a completed diff ($2.00 / $10.00 per 1M tokens) → GPT-6.1 Terra swarms parallel subagents: delivers near-frontier reasoning on tool use at $1.00 / $5.00 per 1M tokens (50% cheaper than 5.6 Terra, 3x faster than Sol) and burns through AST discovery and test suites at 340 tok/s → you stop burning $10/1M Sol context on reading 30 files, cutting agentic swarm costs by 80% while keeping your lead session at a pristine 14k tokens Astra only steps in at three moments: → setting up subagent topology and auth invariant rules before launch → evaluating contract diffs when a test breaks twice or reports an API schema update → signing off on synchronized multi-worktree PRs before merging to main Astra architects · Terra swarms · Sol executes JEV engineering is the same move one layer down: the forks that need no thinker (which worker, which tool, forward or drop) go to JEV in milliseconds, and the big models only see the ones that split the full tree > GPT-6 Astra on call as the adversary auditing subagent topology > GPT-6.1 Terra swarms parallel subagents at $1.00 / $5.00 per 1M tokens with near-frontier accuracy at 340 tok/s > GPT-6.1 Sol patches code and synthesizes PRs when a subagent signals completion > Subagents run in isolated git worktrees with zero conversation history dump > Messages travel as isolated plain text over local IPC with 0 context dump > JEV routes micro-forks in milliseconds paste the tree and this prompt into Codex CLI ↓ "Rebuild my Codex setup around this tree: 1. Configure subagent discovery using ~/.codex/agents and set default_subagent_model to gpt-6.1-terra 2. Delegate all worker execution, AST discovery, and test suites to GPT-6.1 Terra ($1.00 / $5.00 per 1M tokens) at 340 tok/s 3. Spawn GPT-6.1 Sol to execute code edits and synthesize PR merges whenever a subagent reports an update 4. Restrict all subagent traffic to lightweight isolated git worktrees with zero history transfer 5. Reserve GPT-6 Astra for invariant authorization and final sign-off before committing Show me every change as a diff first. No edits until I say go" ↳ http://developers.openai.com/codex/subagents bookmark this, your subagents now swarm without burning lead context
This post quotes another post: x.com/mirku21/status/2098836008468992471