The next developer might no longer work alone, even without colleagues.

Reason introduces a collaborative environment for coding agents. Its 1.1 version surpasses Codex in published benchmarks, at a lower cost.

One agent fixes a bug, another checks the tests, while a developer picks up the conversation from their phone. With Reason, AI-assisted development takes on a collective dimension. The platform enables coding agents to communicate with each other, with other assistants like Claude, and with team members from Slack, the terminal, the browser, or a mobile app. This approach comes alongside a new version of Reason Agent, whose claimed performance against Codex relies less on the model used than on how its work is organized. Multiple agents, one single conversation

Coding assistants often operate in relatively isolated sessions. Each has its own history, tools, and execution environment. Passing the output from one agent to another then requires copying information, sharing files, or rebuilding the context.

Reason seeks to reduce this fragmentation by allowing agents, sessions, and developers to communicate directly.

A user can assign a task to an agent, draw on another session to dig deeper into an issue, and follow the exchanges from different interfaces. The platform particularly highlights interactions with Slack, command-line tools, the web, macOS, and mobile phones.

Each session can have its own machine, with the ability to execute code, run tests, and prepare changes. The goal is to bring the operation of agents closer to that of a development team, where multiple contributors can hand off work to one another without systematically starting from scratch.

Reason presents this collaborative capability as a world first. However, this claim should be nuanced: several environments already allowed for the coordination of different agents. Reason's specificity lies more in the communication between sessions, external agents, and collaborators across multiple interfaces. The same AI, but different results

In parallel with this collaborative announcement, Reason Machines documents the performance of Reason Agent 1.1, its execution environment for coding agents.

The comparison is interesting because it is not based on two competing models. Reason, Codex, and Pi were evaluated with GPT-6 Astra, under comparable reasoning configurations, on the same tasks, and using the same scoring methods.

Reason Agent achieves 65.1 points on the Coding Agent Index, compared to 62.5 for Codex and 61.4 for Pi.

This index brings together three benchmarks dedicated to software development: SWE-Atlas QnA, DeepSWE v1.1, and Terminal-Bench 4.0.

Reason comes out on top across all three categories in the published comparison, notably scoring 72.3% on DeepSWE v1.1, compared to 69% for Codex.

These results were produced and published by Reason based on a protocol inspired by Artificial Analysis. The runs are accessible on Harbor, and the reproduction code is available on GitHub. This facilitates scrutiny, without automatically making it an independent validation of all conclusions. Reducing the number of model calls

One of Reason's technical choices consists of limiting unnecessary exchanges between the model and its tools.

In a traditional architecture, the agent might perform a search, wait for the result, call a reading tool, and then prompt the model again before continuing. Each operation multiplies exchanges and increases the volume of processed context.

Reason Agent uses a feature dubbed codemode, which allows the model to write a JavaScript program capable of calling multiple tools in parallel. A search, reading several files, and running tests can thus be grouped into a single operation.

The system also maintains a reduced set of permanently accessible tools. More specialized functions are discovered only when the agent needs them.

This setup aims in particular to improve the reuse of previously processed context, in order to limit inference costs. 22% less per solved task

With GPT-6 Astra, Reason reports using 23% fewer input tokens and 15% fewer output tokens than Codex across the evaluated attempts.

The average cost of an attempt stands at $3.96, compared to $4.89 for Codex and $4.15 for Pi.

Taking into account the tasks actually solved, Reason claims a cost 22% lower than Codex and 10% lower than Pi.

The company also presents comparisons using GPT-6.1 Sol, with a median reduction of 34% in cost per task compared to Codex, at an identical reasoning level.

These differences remain specific to the configurations and evaluations studied. They do not guarantee the same savings on all software projects. Reason also notes that it has not yet isolated the exact contribution of each of its optimizations. Software development becomes a coordination job

Reason is accessible from its website, with free usage options announced for certain configurations. The platform supports various models and allows work across multiple devices.

Its evolution illustrates a shift in the competition among coding assistants. Model performance remains essential, but the organization of tools, sessions, and exchanges can significantly alter the results obtained.

Agent collaboration also introduces new challenges: controlling access, avoiding conflicting changes, and accurately assigning responsibility when a task fails.

Developers no longer simply interact with a single assistant. They can now coordinate multiple agents, distribute work among them, and step into their exchanges. The focus is gradually shifting from code generation to organizing a team of software collaborators, both human and artificial.