Explore

The product

  • How it worksWhat Trove is, and how your work gets validated and found

Building a Software Factory Powered by Its Own Inference at Subconscious

The Boston Builder Series, from @MGMT Boston in partnership with @Trove.

Dana Wensberg, Head of Platform at Cambridge-based @Subconscious, ships code the way most teams send a Slack message. Tag an agent, review the PR, merge.

Subconscious runs an inference cloud built for long-horizon agents. Earlier this month, the company announced a partnership with Nori, which provides single-tenant cloud environments for coding agents. Nori handles the infrastructure (memory, security, RBAC, integrations, automations), and Subconscious handles inference designed for long sessions. The pitch from CEO Jack O'Brien: hit enter and close your laptop.

The Subconscious team runs its own "product factory" on that stack. Wensberg walked us through two workflows.

1. Bug ticket to PR, no human in the loop

At Subconscious, everything runs through a Slack-based chat app. When a bug or small feature request lands in a support channel, or the team gets off a call with a prospect, someone @-mentions an agent in Nori.

Dana Wensberg's profile
For companiesYour team's validated work, in one place
  • AI visibility reportFree: see whether AI finds you when someone asks
  • The company

    • Our missionWhat we believe
    • AdvisorsWho backs us
    • CareersWork with us
    • ContactGet in touch

    Read

    • BlogEssays and updates
    Explore

    The product

    • How it works
    • For companies
    • AI visibility report

    The company

    • Our mission
    • Advisors
    • Careers
    • Contact

    Read

    • Blog
    Sign inSign up

    From there:

    1. Nori spins up a VM.

    2. It pulls in context from the team's Notion and codebase.

    3. It returns a plan of attack.

    4. It opens a PR.

    5. The team reviews and merges.

    The only human step is the last one.

    2. Let the factory improve itself

    The second workflow is the one Wensberg calls "a really magic experience": telling Nori to improve itself, using the same inference platform it runs on.

    The example is a cron job set up with a single Slack message. Say bugs are piling up and need prioritizing, or you maintain open source repos with a steady stream of outside contributions. One message sets up a job that runs every morning at 9 a.m. and reports the status of every open contribution: which to merge, which to review, which to reject. You check it with your coffee.

    Why long-running matters

    Agents that do meaningful work need to consume a lot of context. Subconscious serves open models tuned for sessions over 200k tokens, including GLM-5.3 Marathon and DeepSeek V4.1 Flash Marathon, which the company says are faster, cheaper and more accurate for long, tool-heavy work. Pair that with Nori's cloud environments and an agent can keep working after you close the tab.

    The takeaway for operators

    Dana's advice applies well beyond Subconscious:

    "Improving and maintaining the factory are just as important as the deliverables it produces. Nori agents powered by our inference make that easy."

    In other words, invest in the factory, not just the output. Make the system easy to tune, and it compounds.

    The secret sauce is making recursive self-improvement easy. Wensberg compares it to tech debt in software engineering: there are always fixes you wish you could make in the moment to improve your own life or your team's, but they take time. A chat interface makes that "gardening" quick. It's Stephen Covey's P (Production) / PC (Production Capability) balance from The 7 Habits of Highly Effective People: don't force the golden goose to lay too many eggs. The deliverables are the eggs; the factory is the goose, and you have to keep it healthy!

    Read the full announcement: Nori and Subconscious Partner on Long-Running AI Agents (Subconscious blog, Oct. 7, 2026).

    The Boston Builder Series highlights people across Boston building with AI, from @MGMT Boston, in partnership with @Trove.

    Media

    • Overview

      Overview

    • Subconscious

      Nori and Subconscious Partner on Long-Running AI Agents

      Nori and Subconscious partner to let teams run long, tool-heavy AI agents in the cloud. Kick off a session in a few clicks and close your laptop.

    Validators

    DW

    Dana Wensberg

    Head of Platform at SubconsciousSubconscious

    SubconsciousSubconscious

    The Boston Builder Series, from @MGMT Boston in partnership with @Trove.

    Dana Wensberg, Head of Platform at Cambridge-based @Subconscious, ships code the way most teams send a Slack message. Tag an agent, review the PR, merge.

    Subconscious runs an inference cloud built for long-horizon agents. Earlier this month, the company announced a partnership with Nori, which provides single-tenant cloud environments for coding agents. Nori handles the infrastructure (memory, security, RBAC, integrations, automations), and Subconscious handles inference designed for long sessions. The pitch from CEO Jack O'Brien: hit enter and close your laptop.

    The Subconscious team runs its own "product factory" on that stack. Wensberg walked us through two workflows.

    1. Bug ticket to PR, no human in the loop

    At Subconscious, everything runs through a Slack-based chat app. When a bug or small feature request lands in a support channel, or the team gets off a call with a prospect, someone @-mentions an agent in Nori.

    From there:

    1. Nori spins up a VM.

    2. It pulls in context from the team's Notion and codebase.

    3. It returns a plan of attack.

    4. It opens a PR.

    5. The team reviews and merges.

    The only human step is the last one.

    2. Let the factory improve itself

    The second workflow is the one Wensberg calls "a really magic experience": telling Nori to improve itself, using the same inference platform it runs on.

    The example is a cron job set up with a single Slack message. Say bugs are piling up and need prioritizing, or you maintain open source repos with a steady stream of outside contributions. One message sets up a job that runs every morning at 9 a.m. and reports the status of every open contribution: which to merge, which to review, which to reject. You check it with your coffee.

    Why long-running matters

    Agents that do meaningful work need to consume a lot of context. Subconscious serves open models tuned for sessions over 200k tokens, including GLM-5.3 Marathon and DeepSeek V4.1 Flash Marathon, which the company says are faster, cheaper and more accurate for long, tool-heavy work. Pair that with Nori's cloud environments and an agent can keep working after you close the tab.

    The takeaway for operators

    Dana's advice applies well beyond Subconscious:

    "Improving and maintaining the factory are just as important as the deliverables it produces. Nori agents powered by our inference make that easy."

    In other words, invest in the factory, not just the output. Make the system easy to tune, and it compounds.

    The secret sauce is making recursive self-improvement easy. Wensberg compares it to tech debt in software engineering: there are always fixes you wish you could make in the moment to improve your own life or your team's, but they take time. A chat interface makes that "gardening" quick. It's Stephen Covey's P (Production) / PC (Production Capability) balance from The 7 Habits of Highly Effective People: don't force the golden goose to lay too many eggs. The deliverables are the eggs; the factory is the goose, and you have to keep it healthy!

    Read the full announcement: Nori and Subconscious Partner on Long-Running AI Agents (Subconscious blog, Oct. 7, 2026).

    The Boston Builder Series highlights people across Boston building with AI, from @MGMT Boston, in partnership with @Trove.

    Validatedby Matt Crane
    M

    Matt Crane

    Founder

    External Collaborator

    Dana is the definition of a builder. Curious, smart, and determined he's been building with AI back when it was called machine learning. This migration of agentic building & finetuning to the chat later should be noticed & followed. Dana...and his approach is the future!

    Related artifacts

    • PSPranav Sonje
      Building A Research & Outreach Agent Workflows With 300+ Nodes
      BlitzyBlitzy · AI Manager, Product Demand & AI Agents
    • BTBrianne Thomas
      How to Run AI Across Your Funnel Without an Engineer
      Elastic PathElastic Path · Head of Marketing / VP Marketing
    • BTBrianne Thomas
      Rebuilt Our GTM Engine So Well That Sales Handed Marketing Its Headcount Spend
      Elastic PathElastic Path · Head of Marketing / VP Marketing
    • BTBrianne Thomas
      Removing the MQL to Increase Pipeline
      Elastic PathElastic Path · Head of Marketing / VP Marketing

    Browse all artifacts →