Deep Agents now swaps in apply_patch the moment you name a Codex model

A profile is selected by the model string. Pass `model="google_genai:gemini-3.1-pro-preview"` into `create_deep_agent` and the matching profile applies with no other configuration ^[12]. What gets applied is more than prompt text. On Codex models the profile replaces the default `edit_file` implementation with `apply_patch` and aliases the `execute` tool as `shell_command` ^[5]. Both names come from OpenAI's Codex Prompting Guide, which LangChain says prescribes tool implementations and names that improve scores on those models ^[6].

So a tool name here is part of the interface the model was tuned against. The prompt edits are the more familiar half. For Codex, LangChain added batching instructions: "Before any tool call, decide ALL files and resources you will need. Batch reads, searches, and other independent operations into parallel tool calls instead of issuing them one at a time." ^[10] For Opus the changes were prompting only, aimed at tool usage and planning, including the line "Active investigation with tools is the default mode of working, not a fallback." ^[11]

The 10 to 20 point figure comes from a curated subset of tau2-bench, which covers multi-turn tool use and instruction following ^[3][4]. LangChain says it picked harder tasks that frontier models have not yet saturated, so that harness-level changes would show up ^[4]. For that margin to transfer, your workload would have to be multi-turn and tool-heavy, and it would have to sit in the band where the model still fails often enough to leave headroom. On tasks a model already passes, a prompt block has nothing to recover. The post gives the range without per-model scores or the size of the subset ^[17].

A second figure from LangChain's earlier harness work is easier to check. It took `gpt-5.2-codex` from 52.8% to 66.5% on Terminal-Bench 2.0 using prompt and middleware-hook changes, moving the submission from Top 30 to Top 5 at the time of publishing ^[8]. The gain is 13.7 points ^[9], inside the range now claimed for profiles, though on a different benchmark and a single model. LangChain also points at the Opus 4.6 submissions on Terminal-Bench 2.0, where the Claude Code harness ranks last ^[7]. The model's own vendor holding the bottom slot for it is the strongest version of the argument.

Automatic application is convenient, and it has a consequence for anyone who reads traces. One source file pointed at two model families now emits different tool names, so a test or a dashboard that asserts on `execute` will not find it under a Codex profile ^[5][12]. Prompt text also becomes a dependency of the package version. LangChain says the default profiles can be inspected in the repo, and that builders can register their own ^[15].

The post's own framing is blunt: "A single harness can't be optimal for every model" ^[13]. Its stated takeaway is narrower than the benchmark number, calling the per-model interface "a helpful primitive for builders to manage profiles per agent, version them, and easily test differences in configurations" ^[16].