LoginSign Up
← Back to all sparks
G

Gumloop

MKT AUTO
Velocity10.0

Gumloop ships Gumball, background subagents, and a model router in two weeks—each one changes how agents work.

ai-agentsmodel-routingautonomous-agentsenterprise-automationself-improving-ai
◆Current state
Gumloop has moved from a workflow automation platform to a multi-agent system with increasingly autonomous capabilities. In early September, the product introduced Gumball—a self-improving, proactive agent running on company infrastructure—background subagents that execute concurrently without blocking the main chat, and a Model Router that picks the right model per task automatically. These are not incremental additions; they represent a shift from 'automations you configure' to 'agents that improve themselves and manage their own execution resources.'
◆Where it's heading
The self-improvement loop is the defining arc: Gumball learns from corrections, agents can now manage their own evaluation criteria, and the Model Router removes model selection as a user responsibility. The daily intelligence layer (Daily Chew, Gumball) is getting tighter with Highlights and actionable briefs. Gumloop is building toward autonomous contributors—proactive, self-assessing, and model-aware—rather than user-triggered task runners.
◆Prediction
The next move is likely deeper agent-to-agent orchestration—routing work between Gumball and specialized subagents based on task type—or expanding the self-improvement loop to cover skill acquisition from past successful runs.

◆Recent moves

  1. 2d ago

    Introducing Gumloop's Model Router

    ⚡ SPARK

    The Model Router is the logical extension of Gumloop's multi-model strategy: instead of users picking the right model per task, Auto mode picks for them. This sits directly on top of the background subagents and Gumball self-improvement capabilities shipped the same week—agents can now run concurrently and route to the optimal model without user intervention.

  2. 7d ago

    Introducing Gumball

    ⚡ SPARK

    Gumball is Gumloop's first agent designed to be proactive rather than reactive: it monitors, surfaces action items via Daily Chew Highlights, learns from corrections, and runs on company cloud infrastructure rather than Gumloop's. This moves the platform beyond 'build and run automations' into 'deploy an autonomous contributor that gets better over time.'

  3. 7d ago

    Update an Artifact From Any Chat

    Artifacts can now be updated across chats and by different agents—keeping a single artifact as the evolving output rather than creating copies—which makes multi-agent workflows coherent around shared outputs. The @gumloop.ai email migration and Slack improvements are operational polish.

  4. 8d ago

    GPT-6 Astra

    GPT-6 Astra joins the Gumloop model roster, and agents can be authorized to manage their own evaluation criteria—editing scoring rubrics or re-running evaluations on past chats with a human confirmation step. The self-evaluation capability extends the Gumball self-improvement arc to any agent on the platform.

  5. 14d ago

    Subagents Run in the Background

    ⚡ SPARK

    Background subagents change the interaction model: agents can delegate to subagents that run concurrently while the user continues the main chat, with per-subagent status and results streaming back as each finishes. Claude Fable 5.1 and Gemini 3.8 Flash ship in the same release, immediately expanding the frontier model options available to all agents.

  6. 16d ago

    A Revamped Connectors Page

    The connectors page is overhauled with category browsing, per-connector tool inspection, and inline account connection. Zip MCP (procurement) and ElevenLabs MCP (audio generation) expand the integration surface. These are operational improvements to connector discovery and management.