Skip to main content
Need help?
Ask in DiscordOpen a GitHub issueSupport guide

​
Summary

This note tracks the current official product signals around deep research as a category, not just as a single case study. It is meant to help contributors extend the lab’s deep-research coverage with current product positioning from OpenAI, Google, and Microsoft.

​
Why It Matters

Deep research is turning into a recognizable agent shape: long-running, source-aware, analyst-style work that gathers evidence, synthesizes it into a report, and increasingly connects to trusted or private systems. That makes it useful across case studies, platform comparisons, and future starter projects.

​
Scope Notes

Included:
  • OpenAI’s current deep research positioning in ChatGPT
  • OpenAI’s current planning, source-control, report-reuse, and workspace signals
  • Google’s developer API framing for Gemini Deep Research
  • Google’s June 2026 Interactions API general-availability shift
  • current Google ADK and A2A bridge signals for Interactions-based agents
  • Microsoft’s migration signal away from its classic Deep Research tool
Excluded:
  • implementation walkthroughs for one specific UI
  • benchmark deep dives beyond what the official product pages mention
  • short-lived quotas, plan limits, region lists, or model-version details
  • broader research-assistant tooling that is not framed as deep research

​
Source Map

  • Introducing deep research: OpenAI describes deep research as a multi-step internet research agent and notes the February 10, 2026 update that added MCP and app connections, trusted-site search controls, and improved progress tracking.
  • Deep research in ChatGPT: OpenAI’s current operator guide describes a reviewable research plan, source selection across sites, files, and connected apps, in-run interruption, cited reports with activity history, and downloads in Markdown, Word, and PDF. It also states that connected apps are read-only during research and documents workspace access controls.
  • Accelerating scientific discovery with ChatGPT for Academic Researchers: OpenAI places expanded deep research alongside larger context, connectors, skills, and Codex in longer scientific workflows. The signal is less about one research feature and more about a governed workspace for research work.
  • Build with Gemini Deep Research: Google exposes Gemini Deep Research through the Interactions API and frames it as an embeddable agent for long-running context gathering and synthesis.
  • Interactions API: our primary interface for Gemini models and agents: Google says the Interactions API reached general availability on June 22, 2026 and now serves as the primary interface for Gemini models and agents.
  • Gemini API docs and Interactions overview: the current docs recommend the Interactions API for new projects and frame server-side state, background execution, and resumable interaction flows as the default developer surface.
  • Building agents with the ADK and the new Interactions API plus the Gemini Enterprise Agent Platform ADK guide: Google now connects the Interactions API to ADK-managed tools, A2A bridges, and enterprise deployment paths instead of treating deep research as an isolated product leaf.
  • Deep Research tool in Microsoft Foundry (classic): Microsoft marks the classic tool as deprecated and recommends using the o3-deep-research model with web search or an MCP tool. This is a useful lifecycle signal: the durable capability can outlast its first packaged tool surface.

​
Synthesis

Five stable signals stand out across these sources:
  • Deep research is being framed as reusable work product, not just chat. OpenAI exposes a report view, citations, activity history, and multiple download formats, while Google frames Deep Research as a long-running synthesis agent.
  • Operator control is becoming part of the product contract. OpenAI now documents reviewable plans, explicit source selection, progress visibility, interruption, and source changes during a run rather than treating the agent as an opaque one-shot request.
  • The product boundary is moving beyond public web search. OpenAI supports files and read-only connected apps plus site controls. Google keeps pushing the same direction through connected data, managed tools, and long-running interaction state.
  • Google is no longer only saying “Gemini Deep Research exists.” The stronger June 2026 signal is that Interactions API is now the primary interface for Gemini models and agents, with Deep Research becoming one concrete surface on top of that wider runtime model.
  • Capability and packaging are separating. Microsoft’s deprecation notice moves developers from a classic bundled tool toward the underlying deep-research model plus web search or MCP. That makes migration paths and lifecycle status part of platform evaluation, alongside quality and feature coverage.
This means the lab should treat deep research as a broader product category with shared traits such as planning, evidence gathering, traceability, operator control, artifact quality, lifecycle management, and long-running interaction state rather than as one isolated feature from one vendor.

​
Contributor Implications

  • Extend Deep Research Agents with a short section that distinguishes end-user deep-research products from the developer-facing runtime surfaces that now expose the same agent shape.
  • Cross-link future updates with Agent Frameworks, Protocols And Interoperability, and the Deep Research Agent Starter so the note points to durable implementation surfaces instead of remaining a pure market snapshot.
  • Compare products and developer surfaces on plan review, source restriction, interruption, evidence visibility, reusable output, connected-data permissions, and migration status instead of comparing only model quality.
  • Route future curation through the Reference Notes hub and the Contributor reading path so product-signal refreshes stay attached to repo-native drafting rules.

​
Gaps And Follow-up

  • Expand Deep Research Agents with a short product-signals section on Gemini Interactions API as a default developer surface for long-running research agents.
  • Add a follow-up note focused on evaluation and trust boundaries for long-running research agents that mix built-in agents with framework-owned outer loops.
  • Add a migration checklist for teams moving between packaged research tools and model-plus-tool compositions without losing source controls, citations, or report traceability.
  • Revisit whether a starter project should support trusted-source constraints or server-side interaction state alongside MCP-connected private data in a future iteration.

​
Update Log

  • 2026-08-11: Added current OpenAI operator and report-reuse controls, the academic-research workspace signal, and Microsoft’s classic-tool migration signal.
  • 2026-06-24: Refreshed the note around Google’s Interactions API general availability and its new role as a primary Gemini agent interface.
  • 2026-04-23: Added a contributor-facing product-signals note based on current OpenAI and Google sources.