New Capability: 41 Expertise Packs — Evidence-Grounded Domain Discipline Across the Agent Org
Date: 2026-05-11 Origin: Agents had codenames, definitions, and Startup Protocols — but the domain knowledge each agent operated from lived in a mix of training data, scattered MEMORY.md lines, and agent-internal informal practice. There was no canonical, queryable, evidence-grounded body of "what does Sentinel actually know about proactive relationship monitoring?" — and no quality bar that distinguished "Sentinel has a definition" from "Sentinel has a discipline." Pack work earlier in 2026 (Spark, Counsel, Scribe, Sage, Trellis, others) established the format; the 2026-05-10/11 stream extended it to the rest of the org. Impact: 41 named agents now operate from evidence-grounded, single-pass, quality-gated expertise packs. Every pack carries forbidden-language discipline, anti-pattern documentation, tooling reference, and (in 39 of 41) verified 2026 URL evidence. The packs are mechanically discoverable (Dewey § 476, INDEX expertise_pack column, V2 manifest) and mechanically loaded (45 agent Startup Protocols reference them by path).
What Was Built
The 41-pack stream added one expertise pack per named agent across V's Operations Org, Sage's Customer Org, Pax's Finance Org, and Cross-Functional + BU Leaders. Each pack is a skills/{pack-domain}/ directory of 6 standard files: a README that frames the discipline; an anti-patterns file enumerating the failure modes; a tooling-2026 file documenting the vendor stack and discipline-vs-marketing critique; and three domain-specific files (varying by agent — typically discipline taxonomy + verification protocol + a craft or operational reference).
Two packs (data-integrity/, service-delivery/) carry 7 files — preserved-plus-six amend pattern from earlier sibling-session work that the 2026-05-10/11 stream brought up to v1.0 evidence-grounded standard.
The 41 conceptual packs ship across 42 directories (Splice + Slate + Vigil share episode-pipeline/; Dub + Caption share transcript-harvesting/). Voice was intentionally skipped per Chris (voice-interface is hardware-bound and has no spawnable agent definition).
Stream Composition
| Wave | Packs | Owner agents | Theme |
|---|---|---|---|
| 6 | 6 | Director, Prelude, Encore, Curator, Splice+Slate+Vigil, Dub+Caption | Media production end-to-end |
| 7 | 7 | Loom, Scrivener, Lookout, Waypoint, Mender, Dean, Chronicle | Platform health + maintenance |
| 8a | 5 | Beacon, Settler, Sync, Catalog, Pixel | Client delivery + content ops |
| 8b | 4 | Audit (amend), Archivist, Klaxon, Tuner | Self-improvement layer |
| 9a | 6 | Sentinel, Prism, Tide, Scout, Quorum, Meridian | Sage relationship intelligence |
| 9b | 4 | Correspondent, Broadcast, Chorus, Steward | Sage content + communications |
| 10 | 5 | Horizon, Atlas, Mint, Tally, Tributary | Pax finance org |
| 11 | 5 | Nexus, Oracle, Relay (amend), Marquee, Exchange | Cross-functional + BU leaders |
| v1.1 grounding | 3 amends | Splice (episode-pipeline), Dub (transcript-harvesting), Prelude (pre-production tooling) | Wave 6 corrective discharge — verified URL grounding |
Total: 41 new packs + 3 v1.1 grounding amends = 44 substantive pack outputs across 45 spawn invocations (Voice skipped).
What "Evidence-Grounded" Means
The Wave 6 corrective established the canonical bar:
"Do not list a source as evidence unless you actually fetched the URL via curl/Bash. If you can't fetch it, flag
[unverified]explicitly."
Three discipline styles emerged across the stream, all three within canon:
- Cite-only-verified (Correspondent, Broadcast, Mender) — drop failed-fetch URLs entirely. Pack contains only URLs the authoring specialist confirmed via HTTP fetch.
- Cite-with-flags (Chorus, Steward, Atlas, Mint, Tally, Tributary, Nexus, Oracle) — keep URL in pack with
[unverified]annotation where the fetch failed. Reader knows the source name; reader knows the URL hasn't been verified in this session. - Cite-with-table (Tuner, Tide, Pixel, Audit, Archivist) — embed a dedicated verification table at the top or end of the tooling file enumerating every URL, fetch result, and HTTP status. Tuner's table covers 36 URLs all HTTP 200; Audit's covers 15 URLs all verified.
All three styles satisfy the audit bar. Q's audit (docs/quality/audits/2026-05-11-pack-stream.md) confirmed no canonical preference — the discipline is "be honest about verification state," not "use style X."
What This Enables
1. Agent Behavior Becomes Inspectable
Before the pack stream, asking "what does Sentinel actually do when it detects a relationship gap?" required reading .claude/agents/sentinel.md Startup Protocol + chasing references to clients/{slug}/sessions/ + interviewing a session that had run Sentinel recently. There was no canonical answer.
After the pack stream, skills/proactive-relationship-monitoring/README.md is the canonical answer. Any contributor (human or AI) can read the pack, audit the discipline, propose changes, or detect drift between the pack and observed agent behavior. The pack is the contract.
2. Discipline Lifts Across Re-spawn
Because the Startup Protocol of each agent now references the pack path explicitly (commit 90e594033), every spawn of the agent loads the pack discipline at session start. Forbidden language, anti-patterns, tooling-2026 vendor critique, and craft references are loaded automatically without per-spawn briefing. Spawn briefs shrink; spawn quality rises.
3. Vendor Critique Becomes Canonical
Each pack's tooling-2026.md file documents the vendor stack the agent's domain operates within — and critiques the vendor framing. For example, signal-progression/tooling-2026.md (Tide) names 6sense, Demandbase, Bombora, ZoomInfo Intent, Clearbit Reveal, and RB2B as the intent-data category, then dismantles the "high fit leads" framing as the velocity-outreach model applied to anonymous traffic. The pack does not ban the vendor; it names how the vendor's vocabulary clashes with Value-First language and what the team translates the vocabulary into.
This is the strongest specimen of the pack format: it teaches the agent how to operate inside a vendor stack without inheriting the vendor's worldview.
4. Cross-Layer Discoverability
Pack discoverability now has three independent surfaces:
| Surface | What it answers | Updated by |
|---|---|---|
.claude/codebase-index.md § 476 |
"What packs exist in the codebase?" | Archivist on pack add/rename/remove |
skills/ai-leadership/INDEX.md § expertise_pack column |
"Which agent uses which pack?" | Aegis on agent roster changes |
V2 skills/_registry/manifest.yaml |
"What packs is V2 aware of for Emerjent runtime?" | V2 pipeline on commit |
A contributor encountering "Tide" can: open INDEX → read expertise_pack: skills/signal-progression/ → cd to the pack → read the README. Zero recall required.
5. V1 ↔ V2 Pack Parity
Pack copies live at both skills/{pack}/ (V1, source of truth) and skills/domains/{pack}/ (V2, copied via commit 356ebf9). V2 AGENT.md Startup Protocols reference skills/domains/{pack}/ paths. When V2 becomes canonical (per memory/reference_v2_platform.md), packs port directly with no rewriting — the language-discipline + methodology-grounded + HubSpot-architecture content does not depend on platform layer.
How to Verify It Works
Verification 1 — Pack file presence
for p in alert-management assessment-to-interest-bridge background-worker-orchestration \
bu-revenue-aggregation capacity-intelligence client-communications client-onboarding \
content-cataloging context-synchronization contributor-intelligence cross-org-convergence \
cta-optimization data-integrity documentation-maintenance episode-pipeline image-generation \
investment-monitoring lms-management media-asset-management media-bu-leadership \
media-monitoring media-post-production media-pre-production media-production \
meeting-documentation memory-curation meta-coaching new-business-intelligence \
office-hours-intelligence proactive-relationship-monitoring renewal-tracking \
revenue-reconciliation self-healing service-delivery signal-progression \
skill-evaluation status-reporting store-bu-leadership team-member-onboarding \
transcript-harvesting website-health-monitoring youtube-community-engagement; do
test -f "skills/$p/README.md" && echo "OK $p" || echo "MISSING $p"
done
Expected: 42 OK lines, 0 MISSING.
Status: VERIFIED 2026-05-11. All 42 pack README.md files present. Pack-file discipline (6 files each, exceptions for data-integrity and service-delivery at 7 files preserved-amend) confirmed in audit docs/quality/audits/2026-05-11-pack-stream.md.
Verification 2 — Agent Startup Protocol references
grep -lE "skills/(alert-management|assessment-to-interest-bridge|background-worker-orchestration|bu-revenue-aggregation|capacity-intelligence|client-communications|client-onboarding|content-cataloging|context-synchronization|contributor-intelligence|cross-org-convergence|cta-optimization|documentation-maintenance|episode-pipeline|image-generation|investment-monitoring|lms-management|media-asset-management|media-bu-leadership|media-monitoring|media-post-production|media-pre-production|media-production|meeting-documentation|memory-curation|meta-coaching|new-business-intelligence|office-hours-intelligence|proactive-relationship-monitoring|renewal-tracking|revenue-reconciliation|self-healing|service-delivery|signal-progression|skill-evaluation|status-reporting|store-bu-leadership|team-member-onboarding|transcript-harvesting|website-health-monitoring|youtube-community-engagement|data-integrity)/" .claude/agents/*.md | wc -l
Expected: 45.
Status: VERIFIED 2026-05-11. Returns 45.
Verification 3 — Forbidden-language compliance
See docs/quality/audits/2026-05-11-pack-stream.md § Forbidden-language scan. Method, raw commands, and results documented there.
Status: VERIFIED 2026-05-11. 286 raw forbidden-token hits across 42 directories, every hit in critique/declaration context. Zero narrative-voice violations.
Verification 4 — Cross-layer wiring (commit SHAs)
| Layer | Commit | Reproduce |
|---|---|---|
| V1 Dewey § 476 | c054f8b61 |
git log --oneline c054f8b61 |
| V1 INDEX.md expertise_pack column | c6b87cbb7 |
git log --oneline c6b87cbb7 |
| V1 45 agent Startup Protocols | 90e594033 |
git log --oneline 90e594033 |
| Pack-authoring-chain closure | 741c90389 |
git log --oneline 741c90389 |
V2 skills/domains/ + manifest |
356ebf9 (V2 repo) |
git -C /mnt/d/Projects/VFT_Platform/2026_VFT_Platform_Infrastructure log --oneline 356ebf9 |
| V2 45 AGENT.md Startup Protocols | 5ae78f5 (V2 repo) |
git -C /mnt/d/Projects/VFT_Platform/2026_VFT_Platform_Infrastructure log --oneline 5ae78f5 |
Status: VERIFIED 2026-05-11. All six commits present and message-matched in audit report.
Risk Tier Classification
Pack Authoring is added to the process register as a new entry:
| Process ID | Process | Owner | Tier | Rationale |
|---|---|---|---|---|
| P-01 | Pack Authoring | Hone (architecture) + Q (audit) + Aegis (batch sweep) + V (orchestration) | 3 (Standard) | Internal infrastructure process; runs on-demand when an agent gets evidence-grounded discipline; no external system writes; no client-facing impact. Self + Q review sufficient. Audit cadence semi-annual. |
Why not Tier 2: Pack authoring does not touch external systems, does not create commitments, does not run frequently enough to warrant quarterly audit. It is an internal capability-construction process.
Why not Tier 1: Pack content errors would not affect production systems, would not affect clients, would not propagate without an obvious failure mode (forbidden-language hits visible in any subsequent spawn).
Why not Tier 4: It is not infrastructure (no service to maintain); it produces canonical artifacts that other processes rely on.
Verification Score
Per docs/quality/verification-protocol.md § Verification Scoring:
Verification Score = (VERIFIED steps / Total steps) * 100%
For Pack Authoring as a process, the operating procedure is the closure-marked skills/apps/pack-authoring-chain.md. The chain has 6 steps. Steps 0 and 6 are closed as obsolete (DEPRECATED per verification protocol). Excluded from denominator. Remaining steps:
| Step | What | Status | Evidence |
|---|---|---|---|
| 1 | Specialist drafts pack files in memory | VERIFIED 2026-05-11 | 41 pack outputs from this stream confirm the pattern; closeout doc enumerates each |
| 2 | Q audits draft for forbidden-language + evidence quality + single-pass discipline | VERIFIED 2026-05-11 | docs/quality/audits/2026-05-11-pack-stream.md is the working evidence |
| 3 | Forbidden-language scan as hard pre-write gate | VERIFIED 2026-05-11 | Stream produced 0 narrative-voice hits across 42 dirs — gate held |
| 4 | Pack files written, agent Startup Protocols patched, INDEX expertise_pack column updated | VERIFIED 2026-05-11 | Commits 90e594033 + c6b87cbb7 (Aegis batch sweep) |
| 5 | git pull --ff-only, commit by explicit path, push to origin/main |
VERIFIED 2026-05-11 | Closeout doc lists commits; explicit-path discipline maintained |
Score: 5 of 5 = 100%. OPERATIONAL.
Decay: Tier 3 process, 365-day decay window. Next re-verification due 2027-05-11 unless a code/data change triggers earlier UNTESTED transition.
Open Items After This Capability
The capability is operational. Two adjacent items remain as separate work:
- Sentinel bundle exceeds Anthropic 100k char cap — opened as CAR
docs/quality/cars/2026-05-11-corrective-sentinel-bundle-cap.md. Sentinel is a Tier 1 cadenced agent; the cap break is a capability regression at the Anthropic Managed Agent surface only (Sentinel remains operational at Emerjent and V1). Does not block the pack stream completion. upload-agents.jsmissing update path — opened as CARdocs/quality/cars/2026-05-11-corrective-upload-agents-missing-update.md. Tier 3 tooling defect. The current script can only create, not update. Worked around in this stream via manual re-upload. Hone has the ticket. Does not block the pack stream completion.- V2 pre-commit hook ENOBUFS on large diffs — opened as CAR
docs/quality/cars/2026-05-11-corrective-v2-precommit-hook-enobufs.md. Platform infrastructure defect at V2 git hooks. Worked around in this stream via--no-verifywith Chris authorization. Does not block the pack stream completion.
Documents in This Capability
| File | Content |
|---|---|
docs/quality/audits/2026-05-11-pack-stream.md |
Audit report (this capability's PASS evidence) |
docs/quality/capabilities/2026-05-11-expertise-packs.md |
This file |
skills/apps/pack-authoring-chain.md |
Pack Authoring chain documentation with closure markers on Steps 0 + 6 |
/mnt/d/AI-WORKSPACE/2026-05-11-pack-stream-closeout.md |
V's stream-execution closeout (metrics, race conditions, sibling-session reconciliation) |
docs/quality/process-register.md § P-01 |
Pack Authoring process register entry (added in this capability work) |
docs/quality/cars/2026-05-11-corrective-sentinel-bundle-cap.md |
Sentinel cap CAR |
docs/quality/cars/2026-05-11-corrective-upload-agents-missing-update.md |
upload-agents.js CAR |
docs/quality/cars/2026-05-11-corrective-v2-precommit-hook-enobufs.md |
V2 hook ENOBUFS CAR |
Reference
- HubSpot Project gating:
553082657648(Expertise Pack Backfill) advances In Development → In Review → Completed on this capability + audit. - Anthropic Managed Agents 100k char cap: encountered on Sentinel only in this stream.
- Pack file format canon: each pack carries 6 files per the established pattern (README + anti-patterns + tooling-2026 + 3 domain-specific).
- V1 ↔ V2 path mapping: V1
skills/{pack}/↔ V2skills/domains/{pack}/.