Skip to content

autor-parity.md

Latest commit

 

History

History
108 lines (87 loc) · 4.83 KB

autor-parity.md

File metadata and controls

108 lines (87 loc) · 4.83 KB

AutoR Parity Map

This note treats AutoR as an upper-layer research application contract and checks whether AgentWorld now exposes the lower-level mechanisms needed to build it without simplification.

Analysis target: AutoR main at 621e7be (Fix Claude session resume recovery).

The goal is not to clone AutoR into AgentWorld. The goal is to make sure AutoR-class systems can be implemented cleanly on top of AgentWorld core.

Design Principle

AutoR mixes three layers:

  1. provider-backed agent execution
  2. durable workflow runtime
  3. research-specific application policy

AgentWorld should own layers 1 and 2 directly, plus the reusable pieces of layer 3 that are still generic enough to belong in a framework.

Parity Table

AutoR design surface AgentWorld mapping Status
durable run directory workspace.RunWorkspace Implemented
run manifest with per-stage state manifest.RunManifest, StageManifestEntry Implemented
stage draft and final files RunWorkspace.stage_draft_path, stage_final_path Implemented
stage attempt state and session continuity operator_state/, stage_session_file, ControllerStageOperator Implemented
fixed stage contracts stage.StageSpec, required_stage_template Implemented
stage markdown validation stage.validate_stage_markdown Implemented
fixed review options and refinement structure stage.FIXED_STAGE_OPTIONS Implemented
repair after invalid or missing draft StageRepairRequest, ControllerStageOperator.repair_stage_summary Implemented
approved cross-stage memory stage.append_approved_stage_summary, rebuild_memory_from_manifest Implemented
stage handoff summaries stage.write_stage_handoff, build_handoff_context Implemented
prior decision ledger context stage.build_decision_ledger_context Implemented
artifact requirements artifacts.ArtifactRequirement Implemented
artifact index with schema inference artifacts.write_artifact_index, format_artifact_index_for_prompt Implemented
literature evidence ledger validation research.validate_literature_evidence Implemented
hypothesis manifest research.write_hypothesis_manifest Implemented
experiment manifest research.write_experiment_manifest Implemented
writing bundle context research.build_writing_manifest Implemented
human approval gate approval.TerminalApprovalGate Implemented
automated validation-only gate approval.AutoApproveGate Implemented
rollback to earlier stage manifest.rollback_to_stage Implemented
resume from existing run root AutoResearchWorkflow.resume(...) Implemented
provider-backed strong-agent operator stage.ControllerStageOperator + controllers Implemented
AutoR-style app factory apps.auto_research.create_auto_research_app Implemented

What Stays Above the Framework

These parts are still application surfaces rather than framework primitives:

  • terminal UI polish
  • web studio / notebook surfaces
  • paper package generation
  • release package generation
  • project bootstrap from an existing repository
  • paper-corpus bootstrap and researcher profile extraction

These are not missing because the core is too weak. They are separate products that can now be built on top of the existing core contracts.

Current Auto-Research Run Layout

An examples/auto-research run now materializes the same class of durable state that AutoR relies on:

run_root/
├── goal.md
├── user_input.txt
├── memory.md
├── run_config.json
├── run_manifest.json
├── artifact_index.json
├── logs.txt
├── events.jsonl
├── logs_raw/
├── prompt_cache/
├── operator_state/
├── stages/
├── handoffs/
└── workspace/
    ├── literature/
    ├── code/
    ├── data/
    ├── results/
    │   └── experiment_manifest.json
    ├── figures/
    ├── writing/
    ├── notes/
    │   └── hypothesis_manifest.json
    ├── reviews/
    ├── artifacts/
    ├── bootstrap/
    └── profile/

Why This Matters

The old simplified version of the auto-research example only proved that AgentWorld could loop over eight stages.

The current version proves something stronger:

  • stage execution is stateful and resumable
  • stage approval is explicit
  • structured artifacts are regenerated and validated during the workflow
  • approved knowledge is promoted into durable cross-stage memory
  • rollback invalidates downstream work instead of silently ignoring it
  • repair is part of the runtime contract, not an afterthought

That is the boundary we need if AgentWorld is going to sit underneath AutoR-class systems instead of being another demo workflow package.