The New Standard Paris: Raycast, Linear and Mistral on building software with agents

Summary

The panel examined how AI agents are changing software development, product design, team structure, and quality control. A recurring theme was that agents make it much easier to build and ship software, but increase the importance of choosing the right problems, preserving product intent, and checking the results.

Models and AI product features

• Mistral had released a new open-weight model at the time of the event. Théo emphasized giving people the freedom to build with it, including in areas such as cybersecurity and finance. He described European model development as a way to provide greater choice and control.
• Linear can be configured to use open-weight models, though its built-in model options and defaults may not immediately include every model.
• Raycast AI supports a range of model providers. Rather than requiring users to choose a model for each task, Raycast uses automatic routing that balances speed, capability, and efficiency. Thomas said model selection involves product judgment, not just benchmark scores.
• Linear uses models for bounded decisions such as classifying incoming issues, assigning labels, and identifying the responsible team. Raycast described using models for classification and safety-related checks, while noting that simpler, faster approaches can be better for some tasks.

Shipping code, reviews, and QA

• The speakers described a shift away from reviewing every line of code and toward checking the result—for example, having an agent exercise a UI and produce a video, or report performance benchmarks.
• Raycast has not required code reviews by default for several years. Its view is that the person shipping a feature remains responsible for whether it works, though sensitive changes, such as database migrations, may merit another person’s review.
• Linear’s view was that agent reviews can help assess risk and may support automatically merging lower-risk work. High-impact areas, such as banking or money transfers, still call for human oversight.
• The panel argued against relying on a separate QA team as the primary quality gate. Raycast and Linear said engineers should test their own work and automate repetitive checks. Raycast also described using agents to test scenarios and report regressions, alongside quick internal feedback after features are shipped.
• Participants cautioned that more generated code can overwhelm reviewers. Proposed responses included better risk assessment, faster automated checks, and reviewing the intent and product rationale before substantial implementation begins.

Building products and workflows for agents

• Linear reported that, during one week, agent-created issues had recently exceeded human-created issues. This was driven in part by integrations that turn conversations into issues. Thomas of Raycast described the countervailing problem: automated workflows can generate so many issues that teams lose sight of what matters.
• Théo argued that products should be designed for agents from the start. That includes documentation and APIs agents can use, MCP support, and planning for identity, permissions, data isolation, governance, and traceability.
• The panel also anticipated shorter feedback loops. In some cases, a team may be able to go directly from a customer report to a fix rather than creating and managing a separate issue.

Craft, teams, and product judgment

• The speakers saw craft and product quality as more important, not less, when implementation becomes cheap. Agents make it easy to ship ideas that have not been thought through, potentially producing bloated or incoherent products.
• Human judgment remains important for deciding what to build, what to reject, and how individual features fit together. The panel stressed the value of a clear product vision, principles, and a record of why past decisions were made.
• Product and engineering roles are becoming more blended. Designers may code, engineers may contribute to design and product decisions, and PMs may build prototypes. The speakers still valued deep expertise, while expecting more people to work across traditional role boundaries.
• Hiring practices are in flux. Linear still uses coding and architecture interviews plus a work trial, but questioned how much a coding interview predicts performance. Mistral discussed the difficulty of finding a better signal and mentioned AI-assisted coding interviews as an emerging approach. Raycast described a relatively compact process focused on collaboration and completing a practical task.
• Engineering leaders said agents make it easier to prototype and stay technically involved. They also warned that prototypes can create extra work for teams if they are handed off without enough polish or context.

Security, trust, and autonomy

• The speakers described using sandboxes, restricted tool access, and network controls to limit agent actions. Raycast said local agents can run in sandboxes where network access may require approval; Mistral discussed sandboxing managed agents and isolating their filesystems. The panel also raised prompt injection and the risk of untrusted content instructing an agent to expose data.
• They compared trust in agents to trust in people: outputs should be checked rather than accepted on authority. Demonstrations, videos, tests, benchmarks, and independent review can increase confidence, but agents can still mislead or make mistakes.
• On autonomy and verifiability, the panel did not offer a single answer. One view was that highly capable agents may remain nondeterministic; another was to combine agents with structured workflows and deterministic components to limit where uncertainty enters.

Other topics raised

• Custom tools and rebuilding products: Participants welcomed rapid prototyping and customization, but noted that reproducing a mature product’s reliability, governance, and scalability is harder than recreating its visible features. They also discussed potential legal questions around commercial copies.
• New devices: The panel saw potential in voice input paired with visual output, augmented reality, and wearable displays, while acknowledging technical challenges such as size and battery life.
• Brand and differentiation: Brand can influence product choice, but the speakers argued that it must be supported by a good product and a consistent record of quality.
• AI economics: Théo argued that open-weight or sovereign models can reduce dependence on providers whose prices or terms may change, giving companies more control over their costs and infrastructure.

Reviews, guides, or tutorials: No formal product review, guide, or tutorial was presented. The session was a panel discussion followed by audience Q&A.

Main speakers and sources

• Thomas H. — Co-founder and CEO of Raycast; opened the event and discussed Raycast AI, development practices, and product decisions.
• Tuomas — Co-founder and CTO of Linear; discussed Linear’s agents, APIs, workflows, hiring, and review practices.
• Théo — VP of Engineering at Mistral; discussed open-weight models, agent security, and building for agents.
• Tanguy — Moderator and Raycast’s first angel investor.
• Audience members — Asked questions on external agents in Linear, AI-assisted development, QA, security, customization, devices, branding, AI economics, and agent autonomy.