Daily Brief

Slash Stack — November 12, 2026

A ranked daily snapshot of AI, software, systems, infrastructure, agents, blockchain, research, and tools from Discover.

· Generated 16:51 UTC

Daily synthesis

Agents move into everyday work, from team chat to phones and decision APIs

Agentes-chat puts people and agents in a self-hosted workspace, while Gemini Decision 1 and Google Research point to other evaluation targets—with important evidence gaps.

Lead analysis

Agentes-chat describes itself as a self-hosted Slack alternative where people and agents share channels and direct messages. Its repository page says agents can run on Claude, ChatGPT/Codex, Gemini or Grok, and lists MCP tools, their own computer and browser, routines, memory, and approvals. It also lists web, desktop, and phone apps, Docker Compose, Postgres, and an AGPL-3.0 license. The practical change is not a new model, but an attempt to place team communication and agents in one workspace that an organization can host itself. That makes it possible to evaluate whether conversations, tools, and approval points work better together than across separate services. The description does not establish the depth of access controls, how agents are isolated, or which integrations work in practice. Because the retrieved repository evidence is truncated, treat these as project-described capabilities, not independently verified behavior. A bounded evaluation could test whether the approval flow and tool boundaries fit real team tasks before sensitive work is delegated. The key unknown is how those advertised controls behave under the team's actual permissions and workflows.

Evidence:[S1] GitHub

Secondary signals

Agentes-chat puts agent collaboration inside a self-hosted team workspace; the practical question is whether its MCP tools and approvals fit everyday work, not whether controls the repository page does not detail can be assumed (S1). Elsewhere, Simon Willison describes Gemini Decision 1 as a decision-model API supporting yes/no, choice, and rubric-scored questions, and notes similarities to Jev. The available evidence is a short excerpt from a discovery source, not primary documentation, so it identifies the proposal without establishing its behavior or limitations (S2).

Google Research describes an investigational framework that turns passive wearable signals into a modifiable “health age” and simulates personalized 12-week habit trajectories with an action-conditioned world model; it also mentions blood biomarkers. The post alone does not establish clinical validation (S4). On mobile, Mika describes an on-device Android agent that operates apps and the web, but the retrieved repository material offers no evaluation results or operational limits (S6). Turri adds an engineering counterpoint: its author says coding agents helped them move faster while leaving them less able to understand their own code. The excerpt does not explain the mechanisms Turri uses to address that problem. Taken together, these projects make useful evaluation questions about not only what agents can do, but where they run and how people retain oversight and understanding (S5).

Evidence:[S1] GitHub · [S2] Simon Willison · [S4] Google Research · [S5] Show HN · [S6] GitHub

Why it matters

  • Agentes-chat is a candidate for evaluating agent collaboration in a self-hosted workspace. Test whether its approvals and MCP tools fit your tasks; the repository page does not establish specific security controls.[S1]
  • Gemini Decision 1 is described with yes/no, choice, and rubric-scoring questions. Willison’s short excerpt is limited; primary documentation and details needed to assess behavior are absent here.[S2]
  • Google Research proposes 12-week habit simulations from wearable signals, but calls the framework investigational. That is not evidence of clinical usefulness.[S4]
  • Mika places an agent on an Android phone to operate apps and the web. It is a mobile evaluation target, though the available project description gives no accuracy, permission, or boundary details.[S6]
  • Turri starts from a concrete coding-agent tradeoff: its author says faster work came with less understanding of the code. The excerpt does not show how the project addresses that problem.[S5]

What to watch next

  • Look for Agentes-chat documentation on agent isolation, MCP permissions, and approvals; those details are absent from the retrieved evidence.[S1]
  • Check for primary Gemini Decision 1 documentation to clarify its decision modes and the comparison with Jev reported by Simon Willison.[S2]
  • Watch for further evidence from Google Research on validation of the investigational framework and its 12-week trajectory simulations.[S4]
  • Examine how Mika behaves when operating Android apps and the web; the available description does not specify its limits or evaluation.[S6]

Automated synthesis · gemini/openai/gpt-6-luna

Source radar

Primary material

Direct links to the sources used in this edition's synthesis.

  1. S1
    agentes-chat/agentes-chat GitHub ·open_source
  2. S2
    Gemini Decisions API Simon Willison ·Article
  3. S3
  4. S4
  5. S5
  6. S6
    wangzan101/Mika GitHub ·open_source
  7. S7
  8. S8
  9. S9
  10. S10
    KKKKhazix/agent-autoloop GitHub ·open_source
  11. S11
  12. S12
  13. S13
    langchain-typesafe==0.1.0 LangChain ·Tool
  14. S14
    KKKKhazix/AIHOT GitHub ·open_source
  15. S15
    llm 0.37.2 Simon Willison ·Article
  16. S16
  17. S17
    libreagi - save OS models Lobsters ·News