Sustained operations · study

An operational role
carried across days.

We studied one coordinator working through scheduled checks, coordination messages, and human prompts over time. The study asks what it takes to maintain continuity—and how to evaluate the cost of doing so.

Study details

What the record shows.

The study reports activity in one snapshot; this page has not independently verified its source records. The figures are not a count of completed tasks or a measure of time saved.

01 / Scope

One coordinator, one snapshot

The record spans roughly twelve days, from September 27 to October 9, 2026. It is a snapshot of one operating role, not an uptime guarantee or a representative sample of every agent.

02 / Recorded activity

25,556 assistant turns

The snapshot contains 25,556 assistant turns and 1,195 scheduled check chains. These are activity counts, not verified completed outcomes.

03 / Method

Study the working record

The study groups recorded turns by trigger, measures usage and timing, and checks a sample of the transition classifier. The study describes the record as a baseline before its proposed interventions.

04 / Interpretation

A quiet check can still matter

A heuristic classified many check chains without a transition. A check can still detect a condition or preserve a deliberate hold, so that estimate is not a measure of wasted work.

05 / Estimate

Repeated context is a model estimate

A reproducible model estimates an upper bound of about 35.6% of recorded cache reads as rereads of injected context, using persistence and characters-per-token assumptions. It is not a measured saving.

06 / Open questions

Outcomes need evidence

Verified completion totals, operator time saved, unattended success, and savings are not yet established. The study covers one role and one window; its classifier sample has uncertainty.

07 / Next extraction

Follow real episodes

The recorded review includes a checkpoint and return associated with the same actor. Continue checking recovery episodes against their source records, operator interventions, and missing evidence.

Inspect the recorded return
08 / Experiment

Change one variable at a time

The study proposes comparing check cadence and working-set changes against independently verified continuity. The study reports these as proposed experiments, not results.

Recorded episode · September 27, 2026

An open question
carried through a return.

A recorded checkpoint and return in the same session as the reconciliation review. This view examines re-entry; it is not a third independent task or outcome.

The responsibility

Keep the pending decision available on return.

Who carries the work?
Checkpoint, session start, and resume records report the same durable actor.
What is being carried?
An identity change was waiting for a decision on how an unknown lifecycle generation should be handled.
Outcome reported by the study
A continuation summary contains the open question and intended next action. A later dispatch reports a revised approach.
Next step reported
Relay the verdict in the continuation summary and dispatch a revised approach; later completion is not established.
What needed a person?

The workflow already needed a decision. The later dispatch reports approval of a revised rule; it does not establish intervention in the context change itself.

  1. 20:24 UTC

    A checkpoint is recorded

    A pre-compaction journal event identifies the coordinator and its bound durable actor.

    Recorded checkpoint
  2. 20:26 UTC

    The actor returns

    Session-start and post-compaction resume events report the same durable actor. The study transcript contains a continuation summary with the pending decision.

    Journal identity fields and captured continuation summary
  3. 20:27 UTC

    The next action is reported

    A dispatch reports approval of a revised approach. No later settlement or equivalent retained context is inferred from these observations.

    Recorded dispatch; chronological association only
Read the recorded account and evidence limits

Availability and verification are reported by the study, not independently verified by this page. References identify records in the source environment; they are not public links or live status.

Who carries the work?

Availability
available
Study-reported verification
reported

The three checkpoint/start/resume records contain the same durable actor ID and report it as bound. This is a recorded identity claim, not authority.

  • nabi://syn/a3d1e8db-285b-49c3-8e35-0215c88c34c5
  • nabi://syn/d6ce4fd1-9dd1-42a6-b6dc-7de9009b278a
  • nabi://syn/39408d59-1758-4ff4-b0df-023a41dce462

What context was selected?

Availability
partial
Study-reported verification
reported

The captured continuation summary contains the open question and intended next action. A ContextPacket for that selection was not located.

Why was it routed to this actor?

Availability
unavailable
Study-reported verification
unverified

A RouteActionRecord for this re-entry was not located.

How does the result connect to prior work?

Availability
partial
Study-reported verification
checked

The source records and matching identity fields were checked. Their ordering does not establish a typed causal chain or equivalent working context.

  • nabi://syn/a3d1e8db-285b-49c3-8e35-0215c88c34c5
  • nabi://syn/39408d59-1758-4ff4-b0df-023a41dce462

Was the event included in an attested batch?

Availability
unavailable
Study-reported verification
unverified

Inclusion verification for the selected re-entry records was not established.

What can happen next?

Availability
partial
Study-reported verification
reported

A later dispatch reports a revised approach. The dispatch lacks a typed actor ID; its association relies on the transcript.

  • nabi://syn/546d5d97-c935-478e-b76c-d35ae8ba18b3

What remains unverified

  • This is another view of the same source session as the reconciliation hold, not an additional completed task.
  • Matching identity and time order do not establish causal linkage, equivalent context, autonomous recovery or a measured recovery duration.
  • The later revised change’s completion and settlement remain unverified.

From research to practice

Start with one responsibility.

A scoped pilot would define the workflow, permitted actions, approval boundaries, budget, and how success will be checked.

Explore the product direction