From Chat to Control Plane - PM Status Report, 4th May 2026

The Week in Summary

A quieter week for model releases, but a significant one for deployment and workflow infrastructure. OpenAI published Symphony, an open-source spec for running agents directly from a project tracker. Mistral shipped Medium 3.5 alongside Work Mode and Vibe Remote Agents - a full agentic workflow stack with connectors for Jira, Linear, GitHub, and Slack. Google’s Gemini added in-chat file generation across a full range of document formats, and an accidental Play Store leak previewed Google’s on-device agent ambitions ahead of I/O on May 19. Anthropic added creative tool connectors across Adobe, Blender, Ableton, Affinity, and Autodesk Fusion. DeepSeek V4, released last week, continued to attract attention for its pricing - V4 Pro undercuts frontier model rates by roughly 78% on document-heavy workloads.

Here’s what shipped.

---

OpenAI - Symphony

Symphony is the most structurally interesting release of the week. It’s an open-source orchestration spec that turns a project management board - built for Linear, but adaptable - into a control plane for coding agents. Every open task gets assigned to an agent. The issue tracker coordinates the work. If an agent stalls or crashes, Symphony restarts it. Teams using it internally reported a 500% increase in landed PRs over the first three weeks. OpenAI released it as a reference spec rather than a maintained product, explicitly inviting developers to build their own implementations on top of it.

---

Anthropic - Creative Connectors

Creative connectors added this week: Adobe Creative Cloud, Blender, Ableton, Affinity, and Autodesk Fusion. These put Claude inside existing creative production pipelines rather than alongside them.

---

Mistral - Medium 3.5 and Work Mode

Mistral Medium 3.5 is a 128B dense model with a 256k context window, available as open weights under a modified MIT licence. Mistral’s data sovereignty positioning is explicit: Medium 3.5 is aimed at finance, healthcare, legal, and industrial engineering teams that need a model they can inspect and host within their own infrastructure.

Vibe Remote Agents let developers launch tasks from the CLI or Le Chat, run them asynchronously on remote cloud infrastructure, and return to finished branches or draft PRs. Multiple sessions can run in parallel.

Work Mode in Le Chat is a new agentic mode for complex multi-step tasks. Connectors for email, calendar, and documents are on by default. Integrations include GitHub, Linear, Jira, Sentry, and Slack or Teams for reporting. Explicit approval is required before sensitive actions execute.

---

Google - File Generation

Gemini now generates finished files directly in-chat: Google Docs, Sheets, Slides, PDF, DOCX, XLSX, CSV, LaTeX, TXT, RTF, and Markdown - exportable and downloadable from within the conversation window.

COSMO, an experimental on-device agent app from Google Research, was briefly and accidentally published to the Play Store before being pulled. Gemini Nano-powered, local/hybrid, with 14 skills including browser automation, deep research, document writing, and calendar management. It previews what Google is building toward for Android ahead of Google I/O on May 19.

---

DeepSeek - V4 and Open-Weight Economics

DeepSeek V4 released last week and now widely available, comes in two variants, both have 1M token context windows. V4 Pro’s pricing is very aggressive and undercuts GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 on price.

---

What This Means for Project Environments

The practical implications this week fall across three areas.

Orchestrated task management is now a deployable pattern. Symphony’s architecture - open tasks mapped to agents, the issue tracker as coordination layer - is directly applicable beyond software development. Any project environment running a structured backlog in Jira, Linear, Asana, or similar has the foundational layer. The tasks that recur week to week (risk reviews, action item follow-ups, change request drafts, dependency checks) are candidates for agent assignment using this model. The spec is open; implementation requires workflow design, not infrastructure build.

In-chat file generation closes a practical gap. Gemini’s ability to produce finished DOCX, XLSX, PDF, and Slides files from within a conversation window - and Mistral Work Mode’s ability to draft outputs across connected tools - removes the copy-paste step between AI output and working deliverable. Status reports, risk register entries, change request templates, meeting minutes: the output can now be a finished file rather than text to reformat. Teams not ready to build orchestrated workflows get immediate value here without any setup.

Open-weight models make a credible case for regulated environments. The combination of DeepSeek V4 Pro’s economics (78% cost reduction on document-heavy workloads) and Mistral Medium 3.5’s open licence and European data boundary positioning means teams in finance, healthcare, legal, infrastructure, and government now have a defensible path to deploying near-frontier capability on private infrastructure. Data sovereignty requirements that previously ruled out enterprise AI licensing are less of a blocker than they were a month ago.

The remaining constraint is workflow design - identifying which tasks are worth building for and defining what a reliable output looks like. That’s a delivery expertise problem, not an AI problem.

What do these updates mean for you?

---

Yes, AI helped me to write this :)