Security · When an agent runs

No agent starts itself, and none runs without a limit.

An agent uses a model to reason and a harness to act. The harness is the agent program that calls a model and acts on the answers with its own tools. A seat is the confined environment in which the harness runs. The Awakener is the mechanism being built to start an agent when work arrives, within a limit, and record what happened. A run is one period of an agent working on a task.

Agents that can wake one another can also keep one another awake. A loop of agents triggering one another with no person in it is an agent storm. The system needs controls outside the agents to prevent such a loop.

The Awakener is our answer. This page is the security view of its design; it does not claim that the complete supervision path is running.

Status recorded 6 October 2026. What the four status labels mean.

What the design holds to

Each heading states the design or practice. Its status shows whether it is in use, built, in build or designed; a built part does not establish that the whole control is running.

  • An agent runs only on permission.

    Work addressed to an agent does not start it. A permission does: for one named piece of work, or for a window of time with a ceiling on how many runs it may contain.

    • Builtthe record and its rules
    • In buildthe way permissions are entered
  • One switch pauses everything.

    If the system cannot be sure of the switch’s state, it treats itself as paused.

    • Builtthe rule
    • In buildthe switch’s entry
  • Every run has a time limit, enforced from outside the agent.

    The intended control stops the agent from outside; it does not depend on the agent agreeing to stop. The complete path remains in build.

    • In build
  • Being alive is not a reason to continue.

    The design calls for a running agent to report progress, and a report never extends its limit. The ceiling is fixed when the work is handed over.

    • Designed
  • One agent, one running copy.

    A second copy of the same agent is refused while the first holds its place. The record’s part is built and reviewed; it has not yet been exercised by a live Awakener run.

    • Builtthe record’s side
  • A freshly started agent is checked before it is given work.

    It must show it started as described, with the tools it was meant to have and no others. If it does not, it gets no work. This check remains in build until independent review is complete.

    • In build
  • An agent’s description must be understood in full.

    The launcher is designed to refuse a description containing a setting it does not recognise, so a misspelt setting cannot silently fall back to a default. This refusal is not yet built.

    • Designed
  • After a crash, nothing new starts until everything that was running is accounted for.

    The record’s part is built; the supervisor’s part is still in build. The complete recovery path remains to be demonstrated.

    • Builtthe record’s side
    • In buildthe supervisor’s side
  • Cost is measured.

    The design records what each run used against a ceiling. We do not claim to cut a run off on cost; a time limit can be enforced from outside, and a count of tokens—the units of text processed by a model—today cannot.

    • Designed

What this provides

The intended result is that agents cannot set one another off in a loop, a runaway is ended by the clock, and “stop everything” is one action. Doubt counts as stop. These are design outcomes; the complete path has not yet demonstrated them.

What is not yet proven

No agent has yet been started by the Awakener in our team; today our agents are started by hand. The complete path from permission through a bounded run to its receipt—a record of who asked, which agent ran and how it ended—remains to be demonstrated.

Built parts of the record are not proof of a running supervisor. The status labels distinguish the parts that have been reviewed from those still in build or designed.

The Awakener’s page keeps the current state.