Request
Classify the work
The system first distinguishes physics continuation, project-system improvement, documentation, validation, or a stop condition.
AI research-agent deep dive / Phase 5A
Research-Agent Workflow
AEther Flow treats research-agent work as a controlled record chain: requests narrow into Director decisions, one bounded AgentJob, completion evidence, and handoffs when continuation is needed. The workflow makes research inspectable; it does not turn process evidence into physics proof.
Workflow lifecycle
A request becomes one inspectable transaction.
The public model is simple: classify the request, inspect source authority, select one route, bind one AgentJob, record completion evidence, and hand off only when the next bounded step is separate.
Decision
Select one route
A Director decision narrows the request to one lane, records the selected path, and preserves rejected alternatives where needed.
AgentJob
Bind the execution
The AgentJob states allowed reads, allowed writes, expected outputs, validators, and claim boundaries for one transaction.
Completion
Record evidence
The completion records changed outputs, command evidence, verdict, uncertainty, and whether a handoff is needed.
Record chain
Each record narrows authority rather than widening it.
Director decisions, AgentJobs, execution-role records, completions, handoffs, and registry rows form a durable evidence chain. Corrections move through superseding packets rather than quiet historical rewrites.
Director Decision Record
Why this route?
Explains the selected role path, rejected alternatives, claim boundary, and checks. It does not create broad future permission.
Execution-role record
Which authority applies?
Binds role semantics to exactly one job. Role templates guide work, while task-local records and allowlists control authority.
Handoff
What remains next?
Names a future route when continuation is needed. A handoff preserves the boundary rather than silently extending the current job.
Registry rows
How is it found?
Make decisions, jobs, roles, and completions discoverable and validator-checkable without becoming scientific proof.
Authority boundaries
Workflow discipline supports research without promoting claims.
Phase 5A explains the operating model only. It does not add routing behavior, alter role authority, or establish physics conclusions.
One bounded job
No hidden second objective
A physics construction should not become benchmark promotion; a documentation repair should not become a schema change.
Source inspection
Memory is navigation
Memory, wiki notes, and generated pages may help locate sources. They do not replace canonical files or registry rows.
Validator evidence
Operational, not scientific
A PASS result can show the transaction satisfied structural checks. It cannot prove a theorem or promote a physics claim.
Human gates
Protected authority stays protected
Ontology adoption, benchmark promotion, Gate Chair decisions, and claim promotion require their own human-gated path.
Reader path
Use this page with source authority and upstream records.
The website can explain the workflow, but the authoritative details remain in tracked source files, registries, schemas, completions, and handoffs in the upstream project.
Back to AI research-agent trackReturn to the track landing page for bounded tasks, role contracts, review discipline, and memory.Source authority pageRead how website pages, generated explainers, registries, and tracked sources relate.Upstream workflow explainerGenerated noncanonical source-backed workflow explainer for public orientation.Director lifecycle explainerGenerated noncanonical guide to Director decisions, AgentJobs, execution-role records, completions, and handoffs.
Source authority
What this page can and cannot establish.
This page can orient readers to bounded research-agent workflow. It cannot change routing behavior, mutate control records, expand a role template into task authority, treat validator success as scientific proof, or replace human-gated decisions.