TechPrufer
Get started
Mistral AI
Mistral AI logoMistral AI · Feature trail

Developer Tooling for Agents

2 posts · Mar 2025 – Sep 2026

What this trail covers

Mistral AI enhances its AI coding assistant, Mistral Vibe, by moving coding agents to the cloud. This enables parallel, asynchronous task execution powered by the new Mistral Medium 3.5 (128B open-weight model) optimized for long-horizon coding. New 'Work mode' in Le Chat facilitates complex, multi-step tasks like research and cross-tool workflows, with human oversight for sensitive actions. The system integrates with GitHub, Linear, Jira, and Slack/Teams. Sessions can be initiated from the CLI. This post details an agentic workflow for automating the creation of Product Requirements Documents (PRDs) and engineering tickets from meeting transcripts, leveraging Mistral Large 2 for advanced natural language processing and structured output generation. The workflow integrates with project management tools like Linear and Jira.

All Mistral AI trails
  1. Modernizing complex legacy code with AI agents | MistralThis post details the application of AI agents for modernizing complex legacy code, specifically migrating 40,000 lines of Fortran 77 to C++. It introduces a structured workflow involving a parity harness for numerical verification, AI agents for documentation generation and code translation, and human oversight for review and unblocking agent stalls. Key technical contributions include the development of a parity harness that instruments Fortran to dump state snapshots and uses a C++ test framework for verification, the use of Vibe CLI and Mistral OCR to document scattered legacy code, and a structured agent workflow (planner, coder, tester, reviewer) with human checkpoints to ensure high-quality, maintainable code. The post highlights the importance of numerical agreement, organized documentation, and structured workflows with human review gates for large-scale legacy migrations.
  2. Version control for prompts & skills in Studio | MistralThis post introduces a system of record for prompts and skills within Mistral Studio, enabling versioning, ownership, and traceability. It addresses the challenge of unmanaged, scattered AI prompts by treating them as production assets with immutable versions, clear ownership, and audit logs. The system allows for fast iteration and controlled deployment, with domain experts and line-of-business owners able to improve production instructions. It integrates with CI/CD pipelines and provides features like rollback, classification labels, and audit logs. The system connects prompts and skills to how the AI actually behaves through Observability, lineage, and telemetry, tracing production output back to the specific asset version.
  3. Remote agents in Vibe. Powered by Mistral Medium 3.5. | Mistral AIIntroduces Mistral Medium 3.5, a 128B dense model with a 256k context window, designed for long-horizon coding and productivity. Enables remote, cloud-based coding agents in Mistral Vibe and Le Chat for parallel, asynchronous task execution. Adds 'Work mode' to Le Chat for multi-step agentic tasks involving cross-tool workflows, research, and synthesis, with explicit user approval for sensitive actions. Allows local CLI sessions to be teleported to the cloud. Mistral Medium 3.5 replaces Devstral 2 in Vibe CLI and becomes the default in Le Chat. The model's reasoning effort is configurable per request. Includes a vision encoder trained from scratch.
  4. Spaces: A CLI Built for Humans and Agents | Mistral AIIntroduces the 'Spaces' CLI, a tool designed for both human developers and coding agents. Key contributions include: 1. Designing CLIs with explicit input methods (flags, defaults) that mirror interactive prompts, enabling programmatic control. 2. Utilizing introspectable plugin systems (defined by data models like BaseModel) as a single source of truth for project components. 3. Generating `context.json` and `AGENTS.md` files to provide agents with structured project snapshots and imperative rules. 4. Enforcing explicit state management by treating CWD, environment variables, and config paths as inputs rather than assumptions. The post frames these principles as essential for building composable, scriptable, and testable developer tools, particularly for agentic workflows.
  5. Rails testing on autopilot: Building an agent that writes what developers won't | Mistral AIThis post details the development of an autonomous agent that generates and improves RSpec tests for Rails applications. It describes the agent's architecture, including its use of Vibe, repository-level context engineering via AGENTS.md, specialized skills files for different Rails file types (models, serializers, controllers, mailers, helpers), and custom tools for RuboCop linting and SimpleCov validation. It also discusses measuring test quality using LLM-as-a-judge and provides concrete examples of RSpec syntax and tool implementation.
  7. Introducing: Devstral 2 and Mistral Vibe CLI. | Mistral AIIntroduces Devstral 2 (123B) and Devstral Small 2 (24B) as new SOTA open coding models. Devstral 2 achieves 72.2% on SWE-bench Verified and is 5x-28x smaller than competitors like DeepSeek V3.2 and Kimi K2. Devstral Small 2 scores 68.0% on SWE-bench Verified and is locally deployable. Both models support a 256K context window. Devstral 2 is designed for production-grade workflows, including codebase exploration, change orchestration, failure detection, and retries. It can be fine-tuned for specific languages or enterprise codebases. Human evaluations show Devstral 2 has a 42.8% win rate against DeepSeek V3.2, though Claude Sonnet 4.5 remains preferred. Introduces Mistral Vibe CLI, an open-source command-line coding assistant powered by Devstral, featuring project-aware context, multi-file orchestration, and IDE integration via the Agent Communication Protocol. Devstral 2 is optimized for data center GPUs (minimum 4 H100-class), while Devstral Small 2 runs on single-GPU or CPU-only configurations.
  8. Announcing Codestral 25.08 and the Complete Mistral Coding Stack for Enterprise | Mistral AIThis post announces Codestral 25.08, which offers a +30% increase in accepted completions, +10% more retained code, and 50% fewer runaway generations compared to prior versions. It also details improvements in chat mode with +5% instruction following and +5% in average MultiplE code abilities. The post introduces Codestral Embed for codebase-scale semantic retrieval, outperforming leading embedding models in code retrieval benchmarks and offering flexible embedding outputs and private deployment. Devstral is presented as an enterprise-ready agentic coding workflow, with Devstral Small 1.1 scoring 53.6% on SWE-Bench Verified and Devstral Medium reaching 61.6%. The Mistral Code IDE plugin is updated to support these capabilities with on-premise deployment options, no mandatory telemetry, SSO, audit logging, and usage observability.
  9. Upgrading agentic coding capabilities with the new Devstral models | Mistral AIIntroduces Devstral Medium and Devstral Small 1.1. Devstral Small 1.1 (24B parameters, Apache 2.0) improves upon its predecessor with a SWE-Bench Verified score of 53.6% and enhanced generalization to prompts and coding environments, supporting Mistral function calling and XML formats. Devstral Medium achieves a SWE-Bench Verified score of 61.6%, is available via API with competitive cost/performance, and supports on-premise deployment and custom fine-tuning.
  10. Introducing Mistral Code | Mistral AIIntroduces Mistral Code, an enterprise-grade AI coding assistant built on the Continue open-source project. It integrates Codestral, Codestral Embed, Devstral, and Mistral Medium models. Key features include an in-IDE assistant for JetBrains and VSCode, local deployment options (serverless, cloud, self-hosted), enterprise tooling with granular admin controls, observability, and SLAs. It enables fine-tuning models on private repositories and supports multi-step tasks like writing modules, updating tests, and executing shell commands under approval workflows. The offering addresses enterprise blockers like limited connectivity to proprietary repos, minimal customization, shallow task coverage, and fragmented SLAs.
  11. Devstral | Mistral AIIntroduces Devstral, an agentic LLM for software engineering tasks, built in collaboration with All Hands AI. Devstral achieves 46.8% on SWE-Bench Verified, outperforming larger open-source and some closed-source models. It is designed to tackle real-world GitHub issues by running over code agent scaffolds like OpenHands or SWE-Agent. The model is versatile, capable of local deployment on hardware like an RTX 4090, enterprise use for privacy-sensitive repositories, and integration into agentic coding IDEs. It is released under Apache 2.0 and available via API as devstral-small-2505.
  12. Empowering product development with an agentic workflow | Mistral AIThis post details the implementation of an agentic workflow for automating the creation of Product Requirements Documents (PRDs) and engineering tickets from meeting transcripts. It leverages Mistral Large 2 for advanced natural language processing and structured output generation, with two key components: PRDAgent and TicketCreationAgent. The workflow is orchestrated to process transcripts, generate PRDs, and then convert PRD content into structured development tickets, which are automatically created in project management tools like Linear/Jira. A Google Colab notebook is provided for implementation.