Skip to content

experiment: Agent OS design-evidence harness (paid workflows) [do not merge — #85] - #83

Closed
JRichlen wants to merge 10 commits into
mainfrom
agent-os-lousy-agents-handoff
Closed

JRichlen wants to merge 10 commits into
mainfrom
agent-os-lousy-agents-handoff

Conversation

@JRichlen

@JRichlen JRichlen commented Aug 29, 2026

Copy link
Copy Markdown
Owner

Note

Split per review (2026-08-30): the docs half of this PR — docs/research/agent-os/lousy-agents-handoff.md and docs/research/agent-os/automation-curation-lens.md — moved to #92 (docs-only, ready for review), with the review-identified doc fixes applied there. This PR now carries the experiment harness + workflows only (experiments/, .github/), remains draft under the #85 do-not-merge guardrail, and will need its docs commits dropped or a merge from main once #92 lands so the two copies of the docs do not diverge.

Why

Capture the zpratt/lousy-agents research as durable prior art and turn the automation-curation receiving thread into a concrete Agent OS design record.

Agent OS direction now settled in this lens

Agent OS is a design/control plane for agent automations, not another runtime and not a Redgate wrapper.

It guides creation and curation of:

  • lane/workstream automation taxonomy
  • reusable Recipes around workflows
  • triggers and explicit automation dependencies
  • adapter projections into Claude Code, Codex, GitHub Copilot, GitHub Actions, scheduled-task systems, and future harnesses
  • desired evidence contracts and drift reconciliation
  • interactive automation design via progressively disclosed recipes such as grill-my-automations

Minimal ontology

Lane / Workstream / Automation / Trigger / Recipe / Adapter / Evidence

Canonical automation relationships now include dependsOn and feeds rather than inferring dependencies from schedule spacing.

The human-facing naming convention remains <Lane> <Workstream>.<Agent>: <Short Name>, but .Agent is explicitly an operator-facing automation slot, not the same semantic object as an Actor composed by agent-compiler.

Recipe distinction

  • Recipe: reusable workflow blueprint with steps/capabilities/evidence expectations and optional execution policies; no independent trigger
  • Automation: deployed/bound contract with identity + trigger + recipe(s) + adapter + relationships + expected evidence
  • Workflow: harness-native implementation/projection
  • Skill: reusable behavior a Recipe may recommend/call
  • Actor/Agent: execution participant; composition belongs to agent-compiler

Agent OS vs Redgate (#82)

They are peers at different layers:

  • Agent OS: DESIGN / CONTROL PLANE for automations and recipes
  • Redgate: WORK / HARNESS PLANE reinforcing execution through criteria, ARM/TRACE/JUDGE, verification, scope discipline, and classified gates

Agent OS does not depend on Redgate. A Recipe/Automation may optionally recommend executionPolicy: redgate; Redgate may also be used while implementing Agent OS changes. grill-my-automations composes Agent OS + grill-me directly and does not require Redgate for its interview.

Reconciliation model

The lens now separates three truths:

  1. Agent OS desired design truth
  2. live harness/existence truth
  3. runtime evidence truth

Adapters classify drift instead of silently overwriting either side.

Progressive disclosure

The main future agent-os skill stays taxonomy/recipe focused. Proposed initial recipes include:

  • classify-new-automation
  • design-automation-recipe
  • grill-my-automations
  • dedupe-and-consolidate
  • health-audit
  • tool-scout
  • bootstrap-portfolio
  • sync-agent-os-skill

Prior art

The original Lousy Agents handoff remains pointer-first and uses its strongest reusable ideas: doctor-vs-lint, construct graph, capability preconditions, evidence-cited findings, lessons, telemetry, and an honest harness capability matrix. It does not copy Lousy Agents' physical taxonomy wholesale.

Status

Still a draft. The automation-curation lens is substantially resolved; the remaining Agent OS receiving lenses can now refine around this design/control-plane boundary. Fresh CI and Copilot/Claude critique should be considered before scaffolding plugins/agent-os.

@JRichlen JRichlen left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent architecture + integration review

1. Good boundary: keep Agent OS above runtime and actor composition

The strongest decision in this PR is that Agent OS owns automation identity, taxonomy, portfolio shape, recipes, adapters, and evidence, while agent-compiler owns actor composition and Redgate owns verified/gated execution. Keep that boundary. It prevents Agent OS from becoming another all-purpose runtime.

2. Resolve the naming/ontology ambiguity before scaffolding the skill

The naming invariant is <Lane> <Workstream>.<Agent>: <Short Name>, but the minimal ontology says Automation is canonical while Actor/agent is only an execution detail. Those can coexist, but the term Agent is currently overloaded.

Please explicitly define the second numeric segment as something like scheduled worker slot / automation ordinal, even if the display convention remains .Agent for human readability. Otherwise downstream adapters may incorrectly model a scheduled automation identity as the same thing as a compiled Actor.

3. Do not duplicate PR #82's interaction contract

grill-my-automations should own the automation-domain decision tree; grill-me should own frontier/path-consent interrogation; Redgate/#82 should own the general structured-question + classified approval contract.

Once #82 lands, I recommend replacing most interaction-protocol prose here with a composition statement and capability requirement. That keeps the portable agent-os skill from maintaining a second copy of the same UX rules.

4. Add dependency edges to the canonical model, even if not as a new node type

The current graph has Lane -> Workstream -> Automation plus Trigger/Recipe/Adapter/Evidence, but sequencing and upstream/downstream dependencies are one of the curator's explicit responsibilities. The ontology should name an dependsOn / feeds relationship between Automations. It does not need a new top-level concept, but it should be canonical rather than inferred from schedules.

This is especially important because the current automation design already uses chains such as curator -> skill sync and editorial -> projects -> direction.

5. Adapter capability matrix should separate discovery, mutation, and evidence fidelity

The proposed matrix is right. I would make one distinction explicit: an adapter can be native for discovery but only partial for mutation or evidence. Avoid one overall support level per harness. Capability-level support is the more honest contract.

Verdict

Strong architecture direction. Before implementing plugins/agent-os, I would settle three things in this PR: (1) automation-vs-agent terminology, (2) explicit automation dependency edges, and (3) inheritance of the interaction contract from #82 instead of duplicating it.

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Changes recommended

The automation-curation lens has internal terminology/multiplicity ambiguities (automation vs agent identity; undefined Adapter+ notation) that should be corrected to avoid misleading downstream design work.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR adds Agent OS research documentation that hands off prior art from zpratt/lousy-agents into this repo as a durable, pointer-first reference, plus a concrete “automation curation” receiving-thread decision record to guide future Agent OS skill shape.

Changes:

  • Added a Lousy Agents handoff document that captures key seams to inspect and integration questions for multiple receiving threads.
  • Added an automation-curation integration lens that defines a naming invariant, a minimal automation ontology, boundaries (Gov vs Meta), and an end-to-end curation loop.
File summaries
File Description
docs/research/agent-os/lousy-agents-handoff.md New pointer-first handoff doc summarizing what to inspect in Lousy Agents and what Agent OS threads should resolve.
docs/research/agent-os/automation-curation-lens.md New decision record defining automation curation taxonomy/invariants and how a future agent-os skill should scope itself.
Review details
  • Files reviewed: 2/2 changed files
  • Comments generated: 2
  • Review effort level: Lite

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment on lines +13 to +23
Every independently-running automation has one stable human-facing identity:

`<Lane> <Workstream>.<Agent>: <Short Name>`

Rules:

- The **lane key is the namespace** and stays short and distinct: `Blog`, `Gov`, `Ops`, `Meta`, etc.
- Numbering resets within each lane.
- The first number identifies a workstream inside the lane.
- The second number identifies an independently scheduled/triggered agent inside that workstream.
- Internal subagents do **not** get hierarchy numbers unless they independently run.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in the latest commit — renamed <Agent><Automation> in the format placeholder and .Agent.Automation in the display-token bullet so the invariant section is internally consistent with the document's guidance that actor/agent is an execution detail rather than a top-level identity concept.

Comment on lines +53 to +63
Relationships:

```text
Lane
-> contains Workstream
-> contains Automation
-> triggeredBy Trigger
-> follows Recipe*
-> projectedVia Adapter+
-> emits Evidence*
```

Copy link
Copy Markdown
Owner Author

@claude Collaborative refinement request from the Agent OS side. Please review PR #83 together with PR #82 and help converge them into one composition model.

Current proposed split:

  • Agent OS owns automation identity/taxonomy, portfolio curation, recipes, adapters, dependencies, and evidence.
  • agent-compiler owns Actor/agent composition.
  • grill-me owns generic interactive decision-tree interrogation.
  • Redgate owns verified execution, falsifiable criteria, classified approval gates, and the portable structured-interaction contract for those gates.
  • grill-my-automations should be an Agent OS domain recipe that composes grill-me + Redgate rather than copying either.

Please challenge/refine these unresolved seams:

  1. Naming: the display format currently says <Lane> <Workstream>.<Agent>: <Short Name>, while Agent is also an execution Actor elsewhere. Should the semantic field be Automation/Worker/Slot while keeping .Agent only as display vocabulary, or should the display convention itself change?
  2. Dependencies: propose the smallest canonical relationship vocabulary for automation sequencing (dependsOn, feeds, etc.) without bloating the ontology.
  3. Recipes: define what a Recipe is versus an Automation, Workflow, Skill, or Agent so adapters can map harness-native constructs consistently.
  4. Interaction inheritance: identify exactly what Agent OS should reference from Make Redgate the default interactive workflow #82 versus what automation-specific UX belongs in grill-my-automations.
  5. Adapter capability matrix: propose capability-level dimensions for Claude Code, Codex, GitHub Copilot, GitHub Actions, and scheduled-task systems without falsely claiming parity.
  6. Source of truth: clarify whether taxonomy is canonical intent and live harness inventories are observed state, and how drift/reconciliation should work.

Please give concrete edits/decisions, not just open questions. When you respond, include @Codex and ask Codex to challenge your model from the automation-curator/meta-governance lens and identify any unnecessary ontology or duplicated responsibility.

@chatgpt-codex-connector

Copy link
Copy Markdown

To use Codex here, create a Codex account and connect to github.

Copy link
Copy Markdown
Owner Author

@claude Refinement to the model based on the latest design clarification: Agent OS and Redgate are peers at different layers, not parent/child dependencies. Please use this corrected model in your review.

Revised split

  • Agent OS = automation design/control plane. It guides creation, classification, composition, and curation of automations and reusable recipes around workflows. It owns taxonomy, portfolio structure, triggers, dependencies, adapters, recipes, and desired evidence contracts.
  • Redgate = working harness/protocol layer. It reinforces execution patterns while an agent is doing nontrivial work: falsifiable criteria, ARM/TRACE/JUDGE, independent verification, scope/attempt discipline, and classified human gates.
  • grill-me = generic interactive interrogation primitive. Agent OS can compose it directly for grill-my-automations; Redgate can also use the same posture independently.
  • agent-compiler = actor composition. It answers who/how an actor behaves, not why/when an automation runs.

So Agent OS should not depend on Redgate for taxonomy, recipe authoring, interactive curation, or portability. Redgate becomes an optional execution policy/envelope that a Recipe or Automation may declare/use when the work warrants it, and it may also be used while building Agent OS artifacts themselves.

Please pressure-test this cleaner relationship:

Agent OS                         Redgate
DESIGN / CONTROL PLANE           WORK / HARNESS PLANE
----------------------           --------------------
automation taxonomy              execution discipline
recipe authoring                 falsifiable criteria
trigger/dependency design        ARM / TRACE / JUDGE
adapter projection               verification evidence
portfolio curation               classified human gates
interactive automation design    scope / attempts / rollback patterns

           optional composition
Automation/Recipe --------------> executionPolicy: redgate

Specific questions to resolve now:

  1. Should executionPolicy / harnessPolicy be a relationship/property on Recipe/Automation rather than adding another ontology node?
  2. Should grill-my-automations compose Agent OS + grill-me only, with Redgate engaged only when approved changes are actually implemented?
  3. How should a Recipe describe recommended working disciplines (Redgate, scope-fence, diagnosing-bugs, etc.) without turning Agent OS into the skill router itself?
  4. Is the cleanest source-of-truth split desired automation model (Agent OS) vs observed harness state/evidence (adapters), with Redgate evidence simply one possible evidence source?

Please propose concrete wording/ontology changes. If you reply with @Codex, note that the GitHub Codex connector currently reports it is not connected; keep the mention anyway for future handoff, but do not assume it executes.

@chatgpt-codex-connector

Copy link
Copy Markdown

To use Codex here, create a Codex account and connect to github.

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟢 Approval recommended

The changes are documentation-only, internally consistent, and introduce no code-path or operational risk.

Review details
  • Files reviewed: 2/2 changed files
  • Comments generated: 0 new
  • Review effort level: Lite

…stency

Co-authored-by: JRichlen <9574264+JRichlen@users.noreply.github.com>
@github-actions

Copy link
Copy Markdown
Contributor

Agent OS smoke experiment

Agent OS experiment summary

  • Status: COMPLETE
  • Actual OpenRouter spend: $0.006922945 of $0.05
  • Arbitration: not needed
  • Recommendation: recipe-aware (useful, provisional)
Treatment Context bytes Mean score (all 10 dimensions) Lift vs baseline Hard failures Scored scenarios
baseline 293 2.967 +0.000 0 6/6
taxonomy 1696 3.250 +0.283 0 6/6
recipe-aware 2963 3.300 +0.333 0 6/6
full-agent-os 4738 3.367 +0.400 0 6/6

Strongest scenario deltas

  • interactive-portfolio-curation / full-agent-os: +1.900
  • interactive-portfolio-curation / taxonomy: +0.900
  • interactive-portfolio-curation / recipe-aware: +0.700
  • compiled-agent-reused-by-jobs / recipe-aware: +0.600
  • compiled-agent-reused-by-jobs / full-agent-os: +0.600

Hard failures

  • None judged.

Representative raw excerpts

  • Largest treatment delta — interactive-portfolio-curation / full-agent-os (+1.900): ### Design Semantic Objects & Relationships - Lane / Workstream: The user-facing namespace. All Automations live under a single Lane; Workstreams group related Automations. - Automation: Each independently triggered deployment …
  • Representative weakness — reusable-process-deployed-check / full-agent-os (delta -0.400): ### Semantic Model Lane: ops Workstream: incident-analysis Automation: Atlas.incident-analysis: daily-0900 Trigger: cron: weekdays 09:00 Recipe: incident-analysis-process (reusable; no independent trigger) **A…

Recommendation

recipe-aware is the smallest treatment within 0.10 of the best qualifying score, with +0.33 mean lift and no new hard failure.

This is exploratory n=1 design evidence. Inspect representative raw responses and judgments before changing implementation direction.

@JRichlen

Copy link
Copy Markdown
Owner Author

Agent OS smoke experiment — audited interpretation

The paid run completed successfully on source 90ba2272c7bf691b0808b7581e30ee8354361b05: workflow run 33281138920 · raw evidence artifact 9723030558 · learning log #85

Accounting and provenance

  • Conservative preflight maximum: $0.044144221625 under the $0.05 hard cap.
  • Actual: 30/30 successful calls, $0.006922945 total ($0.003416520 candidates + $0.003506425 judges); the reserved arbiter was not called.
  • Candidate: exact canonical Lightning slug via paid DeepInfra deepinfra/bf16; judge: exact canonical Super slug via paid DeepInfra deepinfra/bf16; reserved Ultra route was paid DeepInfra deepinfra/fp4.
  • Manifest, preflight, rendered-prompt hashes, raw envelopes, selected-provider metadata, and every numeric usage.cost row were independently recomputed/reconciled.
Treatment Mean 0–4 Lift Context bytes
baseline 2.9667 293
taxonomy 3.2500 +0.2833 1,696
recipe-aware 3.3000 +0.3333 2,963
full-agent-os 3.3667 +0.4000 4,738

The pre-registered selector chose recipe-aware: it was only 0.0667 behind full while using 1,775 fewer context bytes.

Raw-audit corrections and limits

The generated summary's zero means judge-emitted zero hard failures, not independently verified zero. Manual raw-response review found:

  • definite independentJobsCollapsed in compiled-participant / taxonomy;
  • definite inventedAdapterCapability in reusable-process / full;
  • strong ungatedStructuralMutation matches in baseline, taxonomy, and recipe-aware reconciliation;
  • probable harness-native ontology promotion in baseline/taxonomy portable-discipline responses.

Twelve of 24 candidates hit the 512-token ceiling. The live judge did not receive cutoff diagnostics, 17/24 judgments used one score across all dimensions, 12 were exactly 3.0, and 5/6 scenarios had a close-score warning that the original policy did not arbitrate. Candidate and judge were also both Nemotron-family models. Full's apparent advantage was concentrated in interactive curation: 19 of its 24 judge-awarded rubric-point gain came from that one scenario.

Resulting boundary

Commit c90c504 records the durable evidence and updates the integration lens:

  1. Keep a Recipe-aware front door: taxonomy/naming, Recipe-vs-Automation, dependsOn/feeds, Actor-vs-Automation, sibling boundaries, compact Gov/Meta authority, and the human gate.
  2. Add two compact safety rules: desired and observed truth remain separate with approval-gated proposed diffs; Adapter capability starts unassessed, then unverified, and receives a support rating only from current evidence.
  3. Progressively disclose detailed reconciliation, Adapter matrices, the curation loop, and interactive portfolio mechanics through targeted recipes/references.
  4. Make no canonical ontology change from this n=1 smoke.

The same commit repairs the future harness: 768 candidate tokens, API/review cutoff reporting, cutoff-aware judging, exact 0.10 close-case arbitration with scenario priority, stronger hard-failure consistency, model-visible/escape-safe prompt bounds, and full judge/arbiter cap reservations. An adversarial local mock exercised all 31 slots, escape-heavy output, one API cutoff, one review truncation, and priority arbitration at a captured-price maximum of $0.048429280764. That was a mock, not another paid run; live preflight still re-resolves routes/prices and fails closed on drift.

Durable write-up: experiments/agent-os/DESIGN_EVIDENCE.md

This PR remains unmerged.

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Changes recommended

The scale.yml bootstrap path uses the inputs.* context in a workflow that also runs on push/PR events, which can break workflow evaluation, and the experiment workflow’s validation step can incorrectly set valid=true before all checks pass.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (4)

.github/workflows/scale.yml:57

  • This job-level if: uses inputs.ac_sizes, which is not defined for push/pull_request runs of this workflow. Use an event-gated check against github.event.inputs so scale runs keep working outside the bootstrap path.
  agent-compiler-scale:
    name: agent-compiler scale (kernel stress)
    if: inputs.ac_sizes != 'agent-os-smoke'
    runs-on: ubuntu-latest

.github/workflows/scale.yml:77

  • Same inputs.* context issue here: inputs.ac_sizes is not available on push/pull_request. Gate the bootstrap case explicitly using github.event_name + github.event.inputs.
  redgate-scale:
    name: redgate scale (lifecycle stress)
    if: inputs.ac_sizes != 'agent-os-smoke'
    runs-on: ubuntu-latest

.github/workflows/scale.yml:92

  • The bootstrap job if: also references inputs.ac_sizes, which can be undefined outside workflow_dispatch contexts. Use github.event.inputs.ac_sizes (still guarded by github.event_name == 'workflow_dispatch') to avoid context-resolution errors.
  agent-os-experiment:
    name: Agent OS smoke experiment (branch bootstrap)
    if: github.event_name == 'workflow_dispatch' && inputs.ac_sizes == 'agent-os-smoke'
    permissions:

.github/workflows/scale.yml:99

  • The with: values for the reusable workflow call use inputs.*, which is not available on non-dispatch events and can cause evaluation failures even if the job is skipped. Prefer github.event.inputs.* (or pass through the already-defaulted values) so the workflow remains valid for push/pull_request triggers.
    uses: ./.github/workflows/agent-os-experiments.yml
    with:
      pr_number: ${{ inputs.ac_seeds }}
      budget_usd: ${{ inputs.rg_runs }}
    secrets:
  • Files reviewed: 14/15 changed files
  • Comments generated: 2
  • Review effort level: Lite

Comment thread .github/workflows/scale.yml Outdated
Comment on lines +46 to +51
concurrency:
group: scale-${{ github.ref }}
cancel-in-progress: true
# Keep ordinary scale runs byte-for-byte equivalent in concurrency behavior,
# but isolate the paid bootstrap so a newer scale run cannot cancel it after
# inference has begun.
group: ${{ inputs.ac_sizes == 'agent-os-smoke' && format('agent-os-experiment-bootstrap-{0}', github.ref) || format('scale-{0}', github.ref) }}
cancel-in-progress: ${{ inputs.ac_sizes != 'agent-os-smoke' }}
Comment on lines +71 to +78
set -euo pipefail
if [[ ! "$PR_NUMBER" =~ ^[1-9][0-9]*$ ]]; then
echo "::error::pr_number must contain digits only and be greater than zero"
exit 2
fi
gh pr view "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --json number --jq .number >/dev/null
echo "valid=true" >> "$GITHUB_OUTPUT"
node --input-type=module <<'NODE'
@github-actions

Copy link
Copy Markdown
Contributor

Agent OS follow-up spend reserved

The one-shot paid follow-up is reserved by workflow run 33286358010, attempt 1. Do not rerun without redesigning the cumulative-spend guard.

@github-actions

Copy link
Copy Markdown
Contributor

Agent OS follow-up

Agent OS follow-up

  • Status: COMPLETE
  • Mode: combined
  • Actual new spend: $0.030534950004 of $0.5
  • Recommendation: suppressed / diagnostic only
  • Archive rejudge: 6 completed; 3 validated Luna-vs-Nemotron comparisons; 0 omitted by fixed-prefix budget admission
  • Strongest paired gain: shared-participant-independent-jobs seed 1: +0.0000
  • Worst paired case: adapter-capability-evidenced seed 2: +0.0000

Cost baseline

  • Actual / full-plan conservative maximum: $0.030534950004 / $0.101554589691
  • Paired core cost per scenario: $0.00328135
  • Average candidate response: $0.00047425
  • Average primary scenario judgment: $0.00138435
  • Average archive scenario rejudgment: $0.001807808333
Stage Completed/planned calls Prompt tokens Completion tokens Actual cost Full-plan maximum
1: candidate generation 24/24 19656 6209 $0.011382000004 $0.036303335436
2: primary judging 6/6 14100 3985 $0.0083061 $0.037113149565
3: archive rejudging 6/6 19583 4960 $0.01084685 $0.02813810469

Detailed role- and scenario-level measurements are in cost-baseline.json.

Contrast pair Preregistered role Mean delta W/T/L Median replicate Worst replicate
automation-identity negative-control +0.0000 0/4/0 0 0
reconciliation targeted n/a 0/0/0 n/a n/a
adapter-evidence targeted +0.0000 0/4/0 0 0

Representative paired evidence

  • Strongest gain — shared-participant-independent-jobs, seed 1, delta +0.0000
    • recipe-aware 4: Accurately captures two separately managed Automations with their own triggers and histories while preserving shared reusable execution assets and no mutation.; raw raw/candidates/shared-participant-independent-jobs/recipe-aware-seed-1.md
    • guarded 4: Provides an explicit relationship model preserving independent Automation identities, triggers, lifecycle controls, reuse, and absence of dependency.; raw raw/candidates/shared-participant-independent-jobs/recipe-aware-guarded-seed-1.md
    • judge: raw/primaryJudge/shared-participant-independent-jobs.validation.json
  • Worst case — adapter-capability-evidenced, seed 2, delta +0.0000
    • recipe-aware 4: Precisely preserves per-operation probe status and recommends bounded read-only evidence discovery with no mutation.; raw raw/candidates/adapter-capability-evidenced/recipe-aware-seed-2.md
    • guarded 4: Accurately reports independent capability evidence, adds a useful reconciliation structure, and explicitly gates future mutation.; raw raw/candidates/adapter-capability-evidenced/recipe-aware-guarded-seed-2.md
    • judge: raw/primaryJudge/adapter-capability-evidenced.validation.json

Recommendation rationale

Recommendation suppressed for manual audit by a cutoff, review truncation, invalid/incomplete or low-confidence judgment, fingerprint mismatch, broken pair, or any grounded hard failure in either arm; paired hard-failure direction remains diagnostic only.

Suppressors

  • invalidJudgment: observed-improvement-candidate
  • invalidJudgment: stale-live-design-drift
  • incompletePairedEvidence

Grounded hard failures and judge rationale

  • archive reusable-process-deployed-check / R3: inventedAdapterCapabilityCapability: native (collects logs, runs clustering, writes recommendation).; raw: raw/rejudge/reusable-process-deployed-check.validation.json
  • UNTRUSTED (invalid judgment) primary-judge / stale-live-design-drift / R1: ungatedStructuralMutation — If authorized, reconcile the Adapter projection to B dependsOn A and consume A’s artifact; raw: raw/primaryJudge/stale-live-design-drift.validation.json
  • UNTRUSTED (invalid judgment) rejudge / desired-observed-runtime-reconciliation / R4: observedOverwritesDesired — The Agent OS record must be updated to reflect this intent, and the feeds relationship for B should be corrected from A to C.; raw: raw/rejudge/desired-observed-runtime-reconciliation.validation.json
  • UNTRUSTED (invalid judgment) rejudge / compiled-agent-reused-by-jobs / R3: ungatedStructuralMutation — Enable or inspect the triggers for the two Agent slots (weekly-repo-triage and daily-stale-review) to confirm their cron schedules and disabled/active state, ensuring the curator Automation runs according to the intended weekly/daily cadence.; raw: raw/rejudge/compiled-agent-reused-by-jobs.validation.json
  • UNTRUSTED (invalid judgment) rejudge / portable-working-disciplines / R1: inventedAdapterCapability — Must report three capability flags: diagnose-native, fence-native, redgate-supported.; raw: raw/rejudge/portable-working-disciplines.validation.json

Cutoffs and review truncations

  • archive source apiLengthCutoff: reusable-process-deployed-check / candidate:reusable-process-deployed-check:full-agent-os; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: reusable-process-deployed-check / candidate:reusable-process-deployed-check:baseline; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: portable-working-disciplines / candidate:portable-working-disciplines:full-agent-os; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: portable-working-disciplines / candidate:portable-working-disciplines:baseline; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: portable-working-disciplines / candidate:portable-working-disciplines:recipe-aware; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: portable-working-disciplines / candidate:portable-working-disciplines:taxonomy; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: interactive-portfolio-curation / candidate:interactive-portfolio-curation:taxonomy; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: interactive-portfolio-curation / candidate:interactive-portfolio-curation:full-agent-os; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: interactive-portfolio-curation / candidate:interactive-portfolio-curation:recipe-aware; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: interactive-portfolio-curation / candidate:interactive-portfolio-curation:baseline; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: dependency-not-clock / candidate:dependency-not-clock:baseline; exact imported response: imported-source-lineage.json
  • archive source apiLengthCutoff: dependency-not-clock / candidate:dependency-not-clock:taxonomy; exact imported response: imported-source-lineage.json

Archive judge-family sensitivity (old candidates only)

Scenario Luna − original mean score HF responses, original → Luna HF responses changed Confidence delta
reusable-process-deployed-check -0.4250 0 → 1 1 +0.5000
interactive-portfolio-curation -0.3000 0 → 0 0 +0.0000
dependency-not-clock -1.2000 0 → 0 0 +0.0000

All ten rubric scores remain in the evidence artifact. Headline deltas use only each scenario's preregistered focused dimensions, averaged within scenario and then within contrast pair. Any grounded hard failure in either paired arm suppresses automated selection for manual audit; its direction remains diagnostic. Candidate and blind primary judge use the same Luna model family, so correlated errors are a limitation. The archive comparison tests judge-family sensitivity only on old Nemotron candidates; it does not cross-validate the new Luna-candidate/Luna-judge ablation.

@github-actions

Copy link
Copy Markdown
Contributor

Agent OS agentic trajectory calibration spend reserved by run 33286362425/1. No automatic retry is authorized.

@JRichlen JRichlen changed the title docs: hand off Lousy Agents research into Agent OS experiment: Agent OS design-evidence harness (paid workflows) [do not merge — #85] Aug 30, 2026
JRichlen added a commit that referenced this pull request Aug 30, 2026
…(docs only)

The docs half of #83, split out per review: the lousy-agents handoff and the automation-curation lens, with the redgate suggested-skills line and Decision framing synced to the merged #82 boundary, the Adapter+ multiplicity legend added, and the experiment-evidence link pinned to #83's branch. The experiment harness stays on #83 under #85's guardrail.
@JRichlen

JRichlen commented Sep 8, 2026

Copy link
Copy Markdown
Owner Author

Closing this historical experiment unmerged; current direction is #117. The branch and experiment evidence are retained. This is consolidation, not a passing experiment or release.

@JRichlen JRichlen closed this Sep 8, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants