Unruly SystemsResearch & field notes
Systems of work / 2026.09

Commanding an
agentic fleet.

Human attention, correcting direction, and the cost of getting unstuck.

01 / The central findingMake human judgment go further.

A fleet can perform substantial work concurrently while repeatedly returning to one person for interpretation, repair, and decisions. Its practical capacity depends on how that work is organized and how much attention the organization demands from its commander.

“Unblocking agents and correcting their direction.”

Allen’s answer to what currently consumes the most attention — 9 September 2026

This clarification changes the design priority. Better visibility helps identify stalled work. The larger opportunity is to prevent avoidable stalls, detect consequential divergence sooner, and make necessary interventions easier to resolve.

Working thesis

Judge a fleet driver by the verified outcomes it enables per minute of human intervention, while protecting the operator’s ability to understand, redirect, and stop the work.

Deliberate exploration, product judgment, and design discussion remain valuable uses of attention. The target for reduction is avoidable coordination and recovery.

The literature supports treating attention as a scarce organizational resource. It does not supply a universal number of agents that Allen—or any other operator—can safely command. [1] [2]

02 / The cost of getting unstuckClassify the intervention before optimizing it.

“Blocked” combines problems with different owners and remedies. Sending all of them to the human creates unnecessary work; suppressing all of them conceals genuine decisions. The following taxonomy is a design proposal for this fleet.

InterventionTypical triggerBest next move
Recoverable execution failureA failed command, unavailable fixture, or known environment mismatch.Investigate and recover within the assignment’s scope and budget. Escalate with concrete failed evidence when the recovery path is exhausted.
Dependency or capacity holdSeveral candidates need the same GUI session, device, or integration slot.Make the shared constraint visible once, reserve capacity, and stop admitting more work behind it.
Missing knowledgeAn answer exists in project facts, a prior decision, or documentation.Retrieve the answer and cite it. A repeated settled question is a candidate for improving retrieval or the relevant standing instruction.
Direction divergenceWork is internally coherent but pursues the wrong outcome, scope, or quality bar.Expose the interpreted objective and consequential choices early enough to change course. Preserve useful work while reorienting.
Genuine human decisionA novel tradeoff, a change in intent, or a reserved publication or authority boundary.Prepare a bounded decision with evidence, recommendation, consequences, and an explicit response path.

The classifier itself needs evidence. An agent’s confidence that it can proceed is not sufficient authority, and its uncertainty is not sufficient evidence that only a human can help.

Design implication: every intervention should have a reason code and an owner. The driver should report how much demand it resolves, how much it forwards, and which classes recur.

03 / Field notesThe workflow already contains useful lessons.

Operator history

The case material comes from the June 2026 fleet-supervision discussion, a September 9 lane-driver retrospective, the native-driver handoff, and the current coordination documents. These are observed documents and retrospective accounts, not a controlled productivity study.

AttemptObserved lessonConsequence for the design
Terminal mosaic → “Needs You” conceptsThe June work identified the effort of scanning many sessions to discover the one requiring intervention.Make actionable demand easy to find, with native sessions available for detail and control.
Growing coordination narrativesThe September retrospective described large driver documents, competing host views, and expensive reconciliation.Keep one compact current roster. Store history in Git, transcripts, and evidence records.
Parallel implementation lanesSeveral candidates awaited runtime acceptance, evidence transfer, or integration.Reserve acceptance capacity before admitting more authors; prioritize getting existing candidates through their gates.
Short external worker timeoutA worker was given a five-minute outer window; the retrospective acknowledged avoidable startup and commit pressure.Separate the driver’s monitoring period from each worker’s independently bounded assignment.
Cross-host readiness probesA successful probe in a different execution context did not establish that the intended worker session was ready or visible.Verify prerequisites inside the actual shell, with its native identity and attention binding.
Event-driven native-session driverThe current contract includes durable worker identity, worktree ownership, deduplicated wakes, a watchdog, and fixed expiry.Preserve these mechanisms during extraction; evaluate whether they reduce human recovery work.

Current implementation distinction: the inspected wrapper runs normal installed CLIs in their print or exec modes and retains an ordinary shell afterward. Interactive TUI behavior was not established. Its launcher allowlist also lacks the requested yoloastra variant. These are concrete extraction requirements, not reasons to build a new agent runtime.

Historical gate results were not rerun for this research report. The report neither takes ownership of active lanes nor renews their budgets.

04 / Research synthesisSix useful lenses on the same constraint.

Conceptual model + simulations · 2026

Attention demands propagate through a team.

Goodrich, Shields, and Adams model humans, robots, agents, and tasks as a graph of attention demands. Their examples show that different organizational structures can yield different overload and cascading-failure patterns despite identical human–robot ratios. [1]

For this fleet, the application is to examine whether a driver resolves demands or amplifies them through reporting and repeated coordination. The paper’s synchronous assumptions and simplified cognition limit direct predictions for asynchronous coding work. It is a useful design lens, not a calibrated capacity estimator.

Organizational theory · 1971

A supervisor must conserve the recipient’s attention.

Simon treats recipient time as a central cost of information. A subsystem saves organizational attention when it stores or transforms information so that other parts of the organization need to process less of it. [2]

The corresponding driver test is practical: does it settle authorized choices, reconcile evidence, and prepare decisions? A summary that still makes the human reconstruct the situation has conserved little attention.

Economic theory · 2000

Common problems should be solved close to the work.

Garicano’s model explains knowledge-based hierarchies in which common problems are addressed by production workers and exceptional problems reach specialized problem solvers. Communication consumes the solver’s time, including time spent discovering that they cannot help. [3]

Applied here, a valuable intervention can yield a reusable example, decision, or verifier. The benefit depends on future agents finding and applying it. Measure repeat-question rates; avoid assuming that a longer instruction file represents learning.

Human factors analysis · 1983

Recovery requires context that automation can remove.

Bainbridge describes how automation can leave people responsible for difficult exceptions while reducing the routine involvement that maintains their understanding and skills. Effective takeover requires reconstructing the system’s state and its recent behavior. [4]

The design response is an intelligible account of the goal, material changes, uncertainty, and options at the moment of intervention. Native transcripts remain available beneath that account; a quiet queue should still support informed control.

Interaction principles + prototype · 1999

Timing and reversibility influence the cost of help.

Horvitz’s mixed-initiative principles address uncertainty about goals, attention-aware timing, the cost of poor guesses, and efficient user invocation or termination of automated services. [5]

In this setting, urgency should depend on the consequence of waiting and the need for human action. Routine completion, a recoverable failure, and a time-sensitive decision deserve different handling. Explicitly reserved approvals remain explicit.

Human–agent teaming · 2004–2014

Good teammates are understandable and directable.

Klein and colleagues identify common ground, predictability, directability, and coordination costs as requirements for joint activity. Coactive Design makes observability, predictability, and directability central to designing interdependent teams. [6] [7]

A worker should make its interpreted goal and significant commitments inspectable. A correction should update the shared assignment and reach the affected workers. The operator needs evidence that the new direction was understood and applied.

A correction to the June report

The earlier report treated low-single-digit supervisory capacity as a general ceiling. That conclusion is too broad for this fleet. Classical fan-out estimates depend on interaction and neglect time; later work includes queues, reorientation, and situation-awareness delays. None establishes a universal limit for native coding agents. [8] [1]

Measure the frequency, duration, and clustering of this operator’s actual interventions before drawing capacity conclusions.

05 / Proposed operating modelPlace decisions where they can be resolved.

The human retains the mission, consequential tradeoffs, and reserved boundaries. The driver turns that intent into assignments, coordinates dependencies, admits evidence, and resolves interior decisions. Workers execute bounded assignments with their native tools and verifiers.

HUMAN / INTENT & JUDGMENT

Set direction.

Choose the outcomes, explain taste and constraints, resolve novel tradeoffs, and authorize reserved actions.

DRIVER / COORDINATION

Prepare the work.

Translate intent, reserve acceptance, recover within authority, reconcile evidence, and prepare necessary decisions.

WORKERS / EXECUTION

Advance the contract.

Work in native sessions, verify outcomes, checkpoint material changes, and report evidence-backed blockers.

Proposed division of responsibility. Communication can move in both directions; each assignment still has one clear owner.

Correct direction at consequential commitments.

Useful checkpoints occur before work becomes expensive to reverse: choosing an interpretation of the objective, committing to a design, changing a contract, or presenting evidence of completion. They should expose the choice and its consequences. A checkpoint is not automatically a human approval gate.

When Allen corrects direction, the driver should acknowledge the updated objective, identify affected work, revise the authoritative assignment, and record what can be retained. Downstream workers should demonstrate the corrected understanding in their next action or checkpoint.

Convert repeated repairs into durable capability.

For a recurring interruption, determine whether the root cause is missing knowledge, inaccessible context, an inadequate tool surface, an environment defect, or an ambiguous contract. Fix that cause at its owner. Reusable instructions help when judgment is missing; execution failures may need an actual tool or environment repair.

Success criterion: the same class of avoidable interruption becomes less frequent without increasing hidden failures or unauthorized decisions.

06 / A useful interventionBring the human a decision they can make.

An attention request should survive restarts, retain evidence, and distinguish acknowledgment from resolution. It should explain why human involvement is necessary and what work can proceed while the decision waits.

Illustrative request

Resolve a mobile interaction tradeoff

Decision needed
Keep the compact navigation or use the larger, more legible alternative.
Why Allen
Both satisfy the functional requirements. The remaining tradeoff concerns the intended product experience and is not settled in the current brief.
Evidence
Two comparable screenshots, the task results, and the exact candidate revision. Relevant failure cases are included.
Recommendation
Use the larger alternative because the tested primary task requires fewer navigation steps. The compact version preserves more content on the first screen.
Effect of waiting
Visual acceptance stays paused. Independent packaging checks can continue. No release will occur.
Scope
This navigation choice only. It does not approve unrelated layout changes or publication.
Resolution
Record the selected option in the brief, update the worker assignment, and invalidate any screenshots affected by the change.

The same pattern applies to a technical blocker, but the driver should first exhaust the authorized investigative path. A request should include what was tried and why the remaining step needs the human.

Several requests caused by one shared dependency should be represented as one decision with linked consequences. This preserves individual task ownership while reducing repeated context reconstruction.

07 / Attention capacitySmall requests accumulate.

The illustration below calculates the direct service time required by repeated interventions. It is deliberately simple: no queueing model, no empirical prediction, and no claim about Allen’s measured behavior.

48 min / hour
Direct human intervention time

80% of one hour, before additional recovery and switching costs.

workers × requests per hour × minutes per request

Inputs are hypothetical. Above 60 minutes, the requested work exceeds one person’s hour. Below 60 minutes, demand may still be impractical because arrivals cluster, tasks vary, and the operator needs time for other work.

For a real fleet, measure at least three additional costs: time spent recognizing that intervention is needed, time spent reconstructing the situation, and consequential rework after a late correction. A more concise request can reduce service time; clearer assignments and better recovery can reduce request frequency.

Reserve capacity for deliberate thinking. Operating close to a full intervention hour leaves little room for unexpected problems or choosing what should happen next.

08 / Software prior artUseful mechanisms already exist.

These projects were inspected through their published documentation during the research. They were not installed, benchmarked, or compared through controlled use. Their relevance is architectural.

ProjectDocumented mechanismWhat to borrow or examine
Agent DeckSource [10]Persistent conductor sessions, multiple native agent CLIs, child-state notifications, heartbeat fallback, and tool capability differences.The closest comparison for a reusable native-session driver. Examine parent routing, adoption, and explicit launcher capabilities.
Gas TownSource [11]Separate coordinator, worker monitor, and merge-processing roles; durable work tracking.Separate scheduling, worker recovery, and integration when one driver becomes a coordination bottleneck.
Gas Town escalationSource [12]Tracked escalation, severity routing, acknowledgment, stale detection, and bounded re-escalation.Use a durable request lifecycle. Evaluate urgency policy against actual consequence of delay.
Antigravity artifactsSource [13]Plans, code changes, diagrams, and other deliverables support asynchronous review and feedback.Make the work inspectable at meaningful checkpoints. An artifact remains a claim until its supporting evidence is assessed.
Claude SquadSource [14]Multiple native terminal agents in isolated workspaces using tmux and Git worktrees.A simpler baseline for session and worktree management, useful when evaluating the added value of autonomous coordination.

Preserving ordinary installed CLIs, native identity, and resumability is consistent with several of these approaches. The distinguishing work for this fleet lies in acceptance capacity, bounded recovery, and the handling of human judgment.

09 / The reusable driverExtract the contracts that reduce recovery work.

Proposed package boundary

A reusable agent-workflows:driver skill plus a small versioned controller remains a reasonable first package. The skill carries operating judgment. Executable code owns identities, leases, budgets, event delivery, deduplication, receipts, and recovery. Campaign priorities and host reservations remain configuration.

Colocation with the workflow plugin is a provisional starting point. Independent consumers or a separate release lifecycle could justify a separate repository. No extraction or package boundary was ratified by this research.

One assignment should carry enough to act and verify.

  • Objective and permitted scope; repository and worktree; dependencies and acceptance resources.
  • Launcher and its supported fresh-launch, resume, identity, exec, interactive, and event capabilities.
  • Worker identity, host, shell backing, process-start identity, ownership, budget, and terminal conditions.
  • Verifier, evidence identity, and reserved publication or other authority boundaries.
  • Recovery policy and the conditions that warrant a human decision.

Process state, campaign state, and verification state need distinct representations. An exited process can have unfinished work; a running process can already have reached its campaign boundary; a completion claim can still lack admissible verification.

The 2026 Intelligent AI Delegation framework is useful here: it treats delegation as bounded transfer of authority and responsibility and argues that task decomposition should fit available verification capability. It also warns about excessive oversight requests and long delegation chains. This is a proposed framework, not proof of this driver’s effectiveness. [9]

The first acceptance demonstration

Use two ordinary native-CLI workers with bounded assignments. Restart the driver, recover the same workers without duplication, preserve budgets, and consume each completion once. Exercise an execution failure, an unavailable dependency, and a genuine human decision. Confirm that source changes invalidate affected verification evidence.

The operator-facing result should make it possible to understand what changed, what is verified, what remains uncertain, and which decision—if any—requires Allen.

10 / Evaluation proposalMeasure the interruptions that should disappear.

Begin with a small set of comparable real campaigns. Record task mix and difficulty, then compare the current workflow with the proposed driver behavior. This would be an exploratory case study; it would not establish universal effects.

MeasureWhat it revealsGuardrail
Interventions by causeWhich blockers and direction corrections dominate demand.Retain necessary interventions; do not reward suppressing them.
Reorientation timeTime from beginning a request to understanding the decision.Ask the operator or observe directly; transcript gaps are not attention measurements.
Repeated settled questionsWhether prior judgment is found and applied.Distinguish a repeat from a materially changed context.
Correction latency and reworkHow far the fleet proceeds before a wrong direction is corrected.Separate necessary exploration from avoidable divergence.
Verified completion and queue ageWhether acceptance and integration keep up with implementation.Bind completion evidence to the actual candidate and environment.
Protected attentionTime available for intentional design, priorities, and review.Do not equate every minute of human involvement with waste.

Recall can help identify requests, corrections, repeated questions, and session transitions. It cannot by itself measure how long the operator was actively reading or thinking. A short manual observation period or lightweight self-report is needed for that dimension.

Proposed success condition: fewer repeated unblocking requests and less avoidable rework, with equal or better verified outcomes, no increase in missed necessary interventions, and more capacity for deliberate direction.

11 / Limits & open questionsKeep the evidence distinct from the proposal.

  • Transfer across domains: robotics, industrial control, and organizational economics offer useful mechanisms, but their results do not directly establish a coding fleet’s capacity.
  • Evidence quality: this report combines research papers, conceptual frameworks, product documentation, and operator history. These are different forms of evidence.
  • Historical coverage: Recall recovered the June discussion. The September native transcript was read directly after Recall requests timed out. No claim of a complete fleet census is made.
  • Current-state limits: the driver’s documents and wrapper were inspected; its runtime acceptance tests were not rerun for this report. Software alternatives were assessed through documentation only.
  • Unanswered empirical question: the relative contribution of execution failures, missing context, and genuine ambiguity to Allen’s intervention burden remains unmeasured.
  • Design hypothesis: durable decision requests, earlier exposure of consequential choices, and feedback into the right owner may reduce that burden. They still require evaluation.

The next useful investigation is a bounded classification of actual unblocking and direction-correction episodes. It can identify which capability would eliminate the largest recurring demand before the driver grows additional machinery.

12 / Annotated reading listSources to return to.

Start with the attention-demand model, Bainbridge on recovery, and Agent Deck’s conductor. The remaining sources supply organizational, interaction, and implementation context. Links were located during the September 9 research; publisher access policies may vary.

  1. 01

    Reframing the human–robot ratio: Collaborative autonomy and the attention-demand economy

    Michael A. Goodrich, Logan Shields & Julie A. Adams · AI Magazine · 24 August 2026

    A graph-based model of how attention demands propagate and overload teams. Read the limitations alongside the examples: the model is not fitted to this fleet.

  2. 02

    Designing Organizations for an Information-Rich World

    Herbert A. Simon · 1971

    The foundational argument for designing information systems around the recipient’s scarce attention. Particularly relevant to deciding whether another supervisory agent earns its place.

  3. 03

    Hierarchies and the Organization of Knowledge in Production

    Luis Garicano · Journal of Political Economy 108(5) · 2000

    A theoretical account of organizing common and exceptional problem solving when acquiring and communicating knowledge is costly.

  4. 04

    Ironies of Automation

    Lisanne Bainbridge · Automatica 19(6), 775–779 · 1983

    A short, important account of the difficulties left to the operator after automation. Especially useful for thinking about reorientation and intervention support.

  5. 05

    Principles of Mixed-Initiative User Interfaces

    Eric Horvitz · CHI ’99 · 1999

    Principles for combining automated assistance and direct user control, including attention, uncertainty, timing, and recovery from poor guesses.

  6. 06

    Ten Challenges for Making Automation a “Team Player” in Joint Human-Agent Activity

    Gary Klein, David D. Woods, Jeffrey M. Bradshaw, Robert R. Hoffman & Paul J. Feltovich · IEEE Intelligent Systems · 2004

    Common ground, mutual predictability and directability, and the costs of coordination. A useful checklist for what a worker or driver owes its teammates.

  7. 07

    Coactive Design: Designing Support for Interdependence in Joint Activity

    Matthew Johnson and colleagues · Journal of Human-Robot Interaction · 2014

    Designing around interdependence, with observability, predictability, and directability as central requirements.

  8. 08

    Predicting Controller Capacity in Supervisory Control of Multiple UAVs

    M. L. Cummings & P. J. Mitchell · IEEE Transactions on Systems, Man, and Cybernetics—Part A · 2008

    Models waiting, cognitive reorientation, queues, and situation-awareness delays. Useful for understanding why an idealized fan-out estimate is not a universal capacity limit.

  9. 09

    Intelligent AI Delegation

    arXiv:2602.11865 · 2026 · Proposed framework

    Authority, responsibility, capability matching, monitoring, and verifiable completion. Sections on contract-first decomposition and oversight burden are especially relevant.

  10. 10

    Agent Deck

    asheshgoplani/agent-deck · Project documentation inspected September 2026

    Native terminal sessions and persistent conductors, with transition notifications and heartbeat fallback. Inspect capability differences and lifecycle behavior before adopting mechanisms.

  11. 11

    Gas Town

    gastownhall/gastown · Project documentation inspected September 2026

    Coordination, worker monitoring, integration, and durable task tracking are separated into named roles.

  12. 12

    Gas Town Escalation Protocol

    Gas Town design documentation · Inspected September 2026

    Concrete request lifecycle: severity, routing, acknowledgment, stale detection, and bounded re-escalation. Some documented category routing is explicitly future work.

  13. 13

    Google Antigravity: Reviewing Artifacts

    Official product documentation · Inspected September 2026

    Structured deliverables and feedback as a basis for asynchronous collaboration. Artifact presentation and artifact verification remain separate concerns.

  14. 14

    Claude Squad

    smtg-ai/claude-squad · Project documentation inspected September 2026

    A native-agent workspace manager using isolated terminal sessions and Git worktrees. Useful as a simpler comparison point for orchestration overhead.

  15. 15

    Attention-Sensitive Alerting

    Eric Horvitz, Andy Jacobs & David Hovel · Further reading

    Research on the criticality of incoming information and the cost of interrupting its recipient. Located as adjacent reading; not used to estimate this fleet’s performance.

  16. 16

    Display of Information for Time-Critical Decision Making

    Eric Horvitz & Matthew Barry · Further reading

    Methods for choosing what information to show when a person is making a decision under time constraints. An adjacent direction for decision-request presentation.

Case material and provenance

The operator evidence is drawn from unpublished conversations and local coordination artifacts: the June fleet-supervision report and prototypes; the September 9 lane-driver retrospective; the native-driver handoff; and the inspected DRIVER.md, LANES.md, driver contract, and launcher wrapper. Private transcripts, host identifiers, and credentials are not reproduced here.

Design source: Unruly Systems’ HTML effectiveness gallery, using the canonical us-theme v1 semantic tokens and us-brand.js mark engine. This single-file edition embeds its styling, fonts, and scripts; reading does not require external assets.