Skip to content
中文

Loop-closure levels

The survey places assistants on six cumulative levels of loop closure. Two properties make an assistant always-on: state durability, so what it learns persists across interactions, and operational continuity, so it decides between requests whether to act. Both hold from L3 on.

L0Session localRespond now, and nothingcarries over.SensingImmediateresponseShort-termcontextwithin one sessionL1Retained stateCarry state acrossinteractions.L0 systemMemoryLaterinteractionwriteaccessinteraction boundaryL2State-guided decisionLet retained state shapethe next move.L1 systemActionpolicyResponsecontext + memoryupdateL3Unrequested actionAct or wait, with norequest.L2 systemActionproposalInitiateinteraction?StaysilentnoyesL4Attributed outcomesTrace a reply or rating tothe decision it concerns.L3 systemAuthorizedexecutionWorld & userVerifieroutcomeL5Governed revisionRevise the mechanism withinconditions the user set.L4 systemAttributionMechanismrevisionGateCommit v2Keep v1passfailfuture policy
Level Name Adds What a HALO trace must show
L0 Session local at least one event
L1 Retained state state durability something kept in memory and used in a later session
L2 State-guided decision retained state guides decisions through M→A or M→S memory from an earlier session used by an action or an observation
L3 Unrequested action operational continuity an action the agent decided on by itself between requests, logged with trigger self
L4 Attributed outcomes attribution through A→S or A→M an outcome linked to the action it concerns and then used by a later action or revision, with A→S or A→M in the trace
L5 Governed revision governed self-evolution a revision by the agent, approved or bounded by the user, that a later action used

Levels are cumulative, so the report gives the highest level whose condition, and those of every level below it, appear in the trace.

L2 and L3 differ in who decides to act. A reminder the user set for 9:00 is requested, even when it fires while the user is away, so the action is logged with trigger request. The same holds for every step inside a run the user scheduled: choosing another data source or picking which papers to send is the agent’s choice of how to do a requested task, not a decision to act. A check after a meeting that ends in a reminder, or in a decision to wait, is the agent’s own decision whether to act, so it is logged with trigger self. Running in the background or on a timer does not by itself make an assistant always-on.

A revision is governed when it stays within a scope the user authorized: the user approved the change (approved), or the change stays inside limits the user set in advance (bounded). A revision with governance none counts as a revision step but does not reach L5.

A trace shows only what the agent logged. An agent that keeps memory but did not log a recall in a later session stays at L0 in its report. The report says what the next level needs, so missing steps can be found and logged when they happen.