Part IV · Workflows, Planning, Agency, and Authority

Projects, Planning, Branches, and Long-Horizon Alignment

A plan connects a desired outcome to strategy, artifacts, commitments, and the next meaningful move. It stays revisable: evidence and surprises can travel back up the chain and change the route or even the outcome.

PlanPurpose traceabilityBranch budgetCompletion proof

1. A plan is a live theory of completion

AIOS does not define a plan as a task list or calendar. A plan is a revisable account of how a present situation could become a desired outcome and what evidence would show that the transformation occurred. It connects the local choices exposed by menus and workflows to the durable purpose of a project.

Planning is the main bridge through which intent becomes operational without becoming immutable. A project may begin with an exact purpose or with a tension whose implications still need to be clarified. Model judgment can help develop the purpose, compare possible interpretations, and reveal consequences; the person determines which direction becomes governing purpose. Once accepted, that purpose can remain traceable through every planning resolution while results retain the ability to reopen it.

A useful plan connects:

outcome;
strategy;
artifact and workstream architecture;
applicable expertise and accepted best practices;
dependencies and decisions;
bounded commitments;
checkpoints and returns;
foreground movement;
completion evidence;
replan triggers.

2. Five planning resolutions

Five resolutions of a live plan

How a plan translates an outcome into current work while results can move upward to revise every planning resolution.

flowchart TB O["OUTCOME\nFuture condition · beneficiary · scope · evidence"] S["STRATEGY\nRoute · assumptions · alternatives · risk"] A["ARCHITECTURE\nArtifacts · workstreams · decisions · dependencies"] C["COMMITMENTS\nBounded work · authority · checkpoints · returns"] F["FOREGROUND\nNext meaningful movement · why now"] O --> S --> A --> C --> F F -. "results and surprises" .-> C C -. "replanning" .-> A A -. "strategy evidence" .-> S S -. "outcome reflection" .-> O

Plan · canonical15-projects-planning-branches-and-long-horizon-alignment--m01.mmd

Text equivalent

OUTCOME — Future condition · beneficiary · scope · evidence → STRATEGY — Route · assumptions · alternatives · risk; STRATEGY — Route · assumptions · alternatives · risk → ARCHITECTURE — Artifacts · workstreams · decisions · dependencies; ARCHITECTURE — Artifacts · workstreams · decisions · dependencies → COMMITMENTS — Bounded work · authority · checkpoints · returns; COMMITMENTS — Bounded work · authority · checkpoints · returns → FOREGROUND — Next meaningful movement · why now.

The solid descent gives foreground work a traceable reason; the dotted ascent lets surprises challenge commitments, architecture, strategy, and even the outcome. Alignment is therefore maintained through revision, not obedience to the first plan.

Does not establish The resolutions do not collapse into one instruction: strategy does not dictate prose, architecture does not authorize every effect, and foreground does not redefine outcome.

Outcome

Names the future condition, beneficiary, scope, and evidence of completion.

Strategy

Names the governing route, assumptions, major alternatives not chosen, and risks.

Architecture

Names outcomes, artifacts, workstreams, decisions, dependencies, and integration points.

Commitments

Names bounded work, authority, checkpoints, expected returns, and conditions that trigger replanning.

Foreground

Names the next meaningful movement and why it is the right movement now.

Every foreground action should trace upward through commitment, architecture, and strategy to an outcome. Every outcome should trace downward to evidence the work can produce. That path is purpose traceability: the ability to inspect why a local movement exists, which expertise or accepted practice bears on it, what it is meant to change, and where its result will return.

Traceability does not collapse the five resolutions into one instruction. Each resolution translates purpose into the form appropriate to its altitude. A strategy does not dictate prose; an architecture does not authorize every operation; and a foreground movement does not silently redefine the outcome. Results can move upward as evidence, contradiction, or discovery, allowing the person to revise the governing purpose deliberately.

3. Project semantic situation

A project combines:

governing purpose and desired outcome;
purpose lineage and reopen conditions;
current plan across five resolutions;
document and artifact field;
decisions, alternatives, and rationale;
source and evidence field;
applicable expertise, standards, and accepted practices;
active, parked, and completed branches;
assignments, agents, and returns;
relationships and dependencies;
accepted risks and unresolved tensions;
completion criteria and evidence.

The Project Home is a canonical navigational record carrying a compact account of this situation, not a replacement for all underlying files. Canonical artifacts, decisions, and sources remain local and inspectable; indexes and dashboards are derived access structures that can be regenerated.

4. Dependencies are semantic

A dependency exists when later work requires a prior product, decision, acceptance, or condition.

Examples:

a recommendation depends on evidence distinguishing two mechanisms;
a final design depends on an accepted audience definition;
a public document depends on source and quotation review;
integration depends on destination authority;
a workflow stage depends on the semantic product of the previous stage.

“Do A before B” is not enough. The plan should state what B needs from A.

5. Planning and writing are two altitudes

Planning and writing as separate altitudes

How planning commissions structural or writing work, while review returns accepted changes to the plan and structure.

flowchart LR S["Accepted structure"] --> P["Person planning decisions"] P --> O["Structural operations or writing commissions"] O --> W["Writing and production"] W --> H["Holistic review"] H --> R["Refreshed maps, plan, and accepted structure"] R --> P

Semantic altitude15-projects-planning-branches-and-long-horizon-alignment--m02.mmd

Text equivalent

Accepted structure → Person planning decisions; Person planning decisions → Structural operations or writing commissions; Structural operations or writing commissions → Writing and production; Writing and production → Holistic review; Holistic review → Refreshed maps, plan, and accepted structure; Refreshed maps, plan, and accepted structure → Person planning decisions.

Planning and writing meet through commissions rather than collapsing into one act. Production changes the artifact; holistic review decides what holds; refreshed maps and accepted structure then become ground for the person's next planning decision.

Planning can:

change movement order;
create, split, merge, or relate artifacts;
record decisions;
define workstreams and dependencies;
commission writing;
launch a workflow or delegated inquiry.

Planning should not silently write document prose. Writing produces or revises the artifact under the plan's bearing.

6. Branches preserve alternatives and independence

A branch is useful when a line:

requires a different source field;
needs independent or adversarial method;
can produce a useful return without constant parent input;
should remain open while the main line proceeds;
has a bounded receiving decision or artifact;
can mutate a disjoint object.

A branch record should state:

question or alternative;
parent purpose;
reason for independence;
accepted ground and intentional withholding;
source scope;
current state and source snapshot;
expected product and receiving object;
latest useful return;
stop and re-entry conditions;
which unresolved judgments must return to the person.

7. The branch budget

Parallelism has a coherence cost. Before opening branches, the system should consider:

independence gained;
assumptions shared and not shared;
expected disagreement;
integration judgment required;
delay saved versus reconciliation cost;
collision risk;
number of live branches the person can still understand;
whether every branch is waiting on the same unresolved person decision.

The branch budget is a reasoning discipline, not necessarily a numeric limit. More branches are useful only when independence, specialization, or latency savings outweigh duplicated inference, inconsistent assumptions, and reintegration cost.

8. Parallelism requirements

Parallel work is most valuable when:

sources are independent;
genuine specialization matters;
adversarial comparison is useful;
alternatives should not contaminate one another;
mutation surfaces are disjoint;
returns can be compared through a named judgment.

It is harmful when:

purpose or definitions are unstable;
all branches need the same decision;
artifacts collide;
the directing line cannot integrate the returns;
speed is gained only by creating duplicate work.

9. Commitments and authority

A commitment should define:

exact work object;
relationship to parent outcome;
owner or delegated role;
granted context and tools;
mutation ceiling;
expected product;
checkpoint or latest useful return;
completion evidence;
stop and replan conditions;
whether the return is provisional, reversible, and blocked from canonical integration pending review.

A task without a receiving object or completion consequence creates activity without plan intelligence.

A commitment is not identical to the purpose it serves, and neither is identical to authorization. Purpose explains why the work matters. Intentional standing records the state of the decision or commitment. The commitment bounds what should be carried forward. Authorization governs which consequential effects may actually occur. Preserving these distinctions allows a plan to coordinate extensive work without converting broad intent into unrestricted permission.

10. Returns change the plan through judgment

A branch return does not automatically update project truth. It completes the local What of a branch but not the larger Fractal Seed movement; result, reintegration, and renewed purpose occur at the parent altitude.

read return against commission
  → compare with current ground and other returns
    → identify semantic delta
      → integrate into question, plan, artifact, or decision field
        → close, redirect, split, park, or continue

Conflicting returns should be compared by:

source scope;
definitions;
assumptions;
method;
evidence quality;
blind spots;
uncertainty;
time horizon;
factual agreement;
genuine contradiction;
discriminating evidence.

They should not be averaged.

Returns also provide evidence about the intent that commissioned them. A result may satisfy the original purpose, expose an ambiguity in it, reveal that an accepted practice does not fit the case, or show that the desired outcome should change. The model can identify and explain that implication. Only the appropriate person-authority path can revise governing purpose, commitments, or authorization.

11. Long-horizon alignment

Alignment means that present movement, commitments, architecture, strategy, and outcome remain mutually intelligible. It is not obedience to the first plan. It is a maintained, inspectable relationship between current work, governing purpose, applicable expertise, accepted best practices, and the evidence produced as the project develops.

The system can ask:

Does this foreground work still contribute to the outcome?
Has evidence invalidated a strategy assumption?
Is a document being polished after its role in the project changed?
Do current commitments still have authority and resources?
Has an output formed that changes project architecture?
Does a parked branch now satisfy its re-entry condition?
Has the work met stopping conditions even though activity could continue?

Alignment review produces observations and candidate changes. It does not silently rewrite the plan.

12. Replanning triggers

Replanning becomes necessary when:

governing purpose changes;
a critical assumption fails;
contradictory evidence changes confidence;
a dependency cannot be satisfied;
an artifact reveals a different problem;
a branch return changes the decision field;
scope, authority, or resources change;
completion criteria prove inadequate;
residual risk exceeds what was accepted.

The prior plan remains in lineage so the change is explainable.

When a result changes governing purpose, the project should preserve the prior purpose, the evidence or judgment that reopened it, the person-authority decision, the new formulation, and the downstream commitments or artifacts that may now require review. Purpose renewal is therefore a traceable semantic change, not silent drift.

13. Completion proof

Project closure should state:

intended outcome and criteria;
evidence for each criterion;
delivered artifacts and decisions;
unresolved uncertainty or defects;
deferred work and why;
accepted residual risk;
person acceptance where needed;
conditions that would reopen the project.

Files, tasks, model calls, and elapsed time can accompany completion but do not prove it.

14. Example: a source-grounded public guide

Outcome

A coherent public guide that helps a defined audience make a consequential decision without hiding uncertainty or source limitations.

Strategy

Define the decision and audience first, assemble the evidence that could change the decision, separate source findings from interpretation, and review the final claims against the original material.

Architecture

Decision frame; source set; evidence map; draft; independent source review; publication and update path.

Commitments

Give each section a distinct purpose, preserve source descent, record unresolved questions, and require an independent claims review before publication.

Foreground

Resolve the audience and decision criteria before drafting the recommendation.

Replan trigger

If strong evidence changes the decision frame or reveals that two sources use a key term differently, revise the guide's structure and preserve why the plan changed.

This demonstrates planning as a theory of completion rather than a checklist of filenames.

15. Failure modes

Task-list planning

Activities lack outcomes, dependencies, and evidence.

Foreground drift

The next action is locally attractive but no longer serves the project.

Parallelism theater

Many branches create duplicated work and reintegration burden.

Plan rigidity

Past commitments remain active after their assumptions fail.

Return dumping

Delegated products accumulate without a comparison and integration judgment.

Completion by exhaustion

The project closes because time or attention ended rather than because criteria were satisfied.

Hidden authority

A planner or agent silently changes scope, purpose, or commitments.

16. Research evaluation

Evaluate:

traceability from foreground work to outcome;
traceability from local judgments and operations to governing purpose;
alignment with applicable expertise and accepted best practices, including justified departures;
quality of dependency descriptions;
speed and fidelity of replanning after contradiction;
usefulness of branch-budget decisions;
reintegration quality for conflicting returns;
reduction in abandoned or duplicate work;
completion evidence versus activity metrics;
plan recovery after interruption;
whether planning altitude improves project coherence without constraining formation;
correction and regression rates after branch returns;
human review burden created by provisional asynchronous work.

17. Research connection: durable planning answers a real reliability problem

AIOS proposes the live plan, bounded branch, and explicit return contract as the way a reasoning environment remains coherent across time. Current agent research supports the problem this architecture addresses. τ-bench shows that valid tool use and plausible trajectories do not ensure correct policy-and-state outcomes or stable repeated performance. FeatureBench finds that agents that perform well on bounded repair can remain weak on larger multi-file feature work.

Those results support exact state checks, bounded commitments, and persistent plans as serious design requirements. They do not establish that the five AIOS planning resolutions are optimal or that a durable plan automatically improves a model. The relevant comparison is longitudinal: whether the plan helps a person and model resume, redirect, integrate, and prove completion better than a transcript, task list, or broad autonomous objective.

Read deeper in the AI-native architecture brief and its full research memo.

The plan determines what may continue without the person's immediate attention. The next chapter develops that temporal relationship as attention-following asynchronous work.