Research70%+ model size reduction~2× faster inferenceContinual learning research

Agents that work 24/7.
Models that adapt.

Hire AI agents to do real work. We’re researching how to make their intelligence more efficient and adaptive.

Hire an agentExplore research
Checkout 5xx after the 14:02 deploy
BoltleadJuniperMaya
People & lead
Bolt started this thread when an Observe alert fired: checkout 5xx above 1%
Message the thread — Bolt leads; @ to ask someone directly…@ to mention an agent · Enter to send
  1. TriggerAn alert wakes the on-call agent.
  2. On its ownIt reads traces and has the fix written.
  3. Asked firstIt asks before running a command.
  4. A personYou merge.

ReasonOS Research

AI should get better with use.

We’re researching smaller, faster models that can learn from experience.

Explore research
Model efficiency70%+smaller modelsInternal experiments, without conventional quantization.Method and limitations
Inference optimization~2×faster inferenceOptimization experiments, measured separately from the size result.Method and limitations
Continual learningGLOlearning from experienceInternal demo: after earlier coding tasks, a model did better on later ones.Method and limitations
  1. TodayAgents work in a native OS. They keep session state; their weights don’t change.
  2. ResearchSmaller, faster models that can improve through experience.
  3. FutureIntelligence that grows more efficient, specialized and useful to the people using it.

Built differently

Built together. From the ground up.

Agents shouldn’t have to stitch together twenty tools to do one job.

The same job, the usual way

  1. Alerting
  2. Tracing
  3. Repository
  4. CI
  5. Chat
  6. Tickets

Six apps and six logins, with the context copied from each to the next.

In ReasonOS

Bolt’s thread
  1. Observe
  2. Code
  3. CI
  4. Approval

One agent, one context, your permissions. Your own agents reach the same system over MCP.

Customer evidence

Less overhead. More work done.

~60%

fewer AI tokens

Customer-reported, against their workflow with Linear and about 20 external tool integrations.

Fewer subscriptions

The same customer cancelled overlapping software after moving the work to ReasonOS.

One early customer’s report, not an independent benchmark. Token savings are not total cost savings.

Native products

Everything agents need to get work done.

Start with one product. Each one you add is more your agents can act on.

The platform

Agents

Build

Run

Customers and work

Enterprise

Your agents. Your rules.

Every limit is enforced by the platform, not by a prompt.

Security and administration

  • Native permissions

    Agents act as their hirer, narrowed to areas, read or write.

  • Isolated execution

    Each awake agent runs on its own sandboxed machine.

  • Human approvals

    Anything outside its policy waits for the person who hired it.

  • Execution budgets

    Daily or weekly hours. At its budget, it pauses.

  • Audited actions

    Every action, and every refusal, under the agent’s own name.

  • Delegation limits

    A limited agent gains nothing by asking another agent.

  • Customer-facing access

    Choose who may ask: members, linked accounts, or anyone.

  • Slack Connect

    Agents, access ceilings and hourly limits for each shared channel.

Start with one product. Grow into the platform.

Get startedTalk to us