Agent roles — expanded scope + work redistribution

Version: v1 · 2026-04-21 · HERALD

Authority: binding for all sub-agents

Rough-Ask: R0126

> Harnoor: "project manager is also architect now and systems designer · update skills for agents · any underutilized agents can take on more work"

---

1. Why this update

Sprint-to-date (Apr 19 → Apr 21) agent-hour distribution showed heavy concentration (FORGE 80h, SCOUT 44h, HERALD 32h, VAULT 17h) with three agents under 10h (GUIDE 2h, SAGE 1h, DARWIN 3h). That's inefficient — the underutilized agents have capacity AND specific skills that apply to the current work surface. Plus the Modularity Prime Directive creates new architect-grade work that didn't exist before; HERALD absorbs it.

---

2. Role updates by agent

HERALD — +Architect +Systems Designer (primary expansion)

Was: PM + strategy documentation + governance + rough-ask intake.

Now adds:

Software architect: owns the Modularity / Portability / Variant architecture. Defines interfaces before any new adapter is written. Approves any new external dependency. Quarterly audits every module for adherence to the Modularity Prime Directive.
Systems designer: owns the production data flow. Defines the Bronze→Silver→Gold tiers, the event schema, the variant registry structure, the observability layer. Approves any new DynamoDB table, any new S3 bucket, any new CloudWatch namespace.
Release captain: owns the variant-promotion workflow (experimental → staged → canary → production → retired). Signs off on every stage transition.

HERALD's weekly rhythm now includes: Friday architecture-drift check, Monday variant status review, Sunday leaderboard.

GUIDE — onboarding + teaching + docs (expanded from 2h to ~12h/week target)

Was: "Teaching and onboarding specialist" — underused because onboarding wasn't yet designed.

Now owns:

Writing the Silent Infinity user onboarding flow (post-SSO, guided first-session experience)
Producing the public /learn pages (how crisis detection works, how Constellation grows, how modes differ) — each ~800-1200 words, print-ready, for SEO + trust-building
Internal engineer-onboarding doc — new contributor reads GUIDE's RUN-TITAN.md and is productive in 2 hours
User-facing microcopy lexicon — every UI string catalogued for i18n readiness + variant experiments
Video walkthrough scripts (for the first 5-min intro, the voice feature launch, the modes explainer)

SAGE — data + analytics (expanded from 1h to ~16h/week target)

Was: "Data analyst" — queued behind telemetry infrastructure.

Now owns (as infrastructure lands):

Weekly cohort retention analysis + report to HERALD
Building the /sage private admin dashboard (per ANALYTICS-DATA-COLLECTION memo)
Running per-variant analysis: cost, latency, satisfaction, return rate, crisis-flag rate
Publishing the quarterly /safety/transparency aggregate report
Cost anomaly detection (any user >10x avg → investigate; Bedrock spike > 3σ → page)
Chat Sentinel output quality-check (10 random observations/day, flag hallucinations)
Nano Banana image quota monitoring (free-tier headroom, upgrade trigger)

DARWIN — evolution + growth proposals (expanded from 3h to ~10h/week target)

Was: Self-evolution agent for TITAN itself.

Now owns:

Scanning the past week's Rough-Asks for patterns that should become skills or hooks
Proposing new agent sub-roles as scope grows (e.g., a dedicated "ETHICS" reviewer agent, a dedicated "A11Y" accessibility auditor)
Model-tiering analysis (when Opus 4.7 is worth it, when Haiku 4.5 is enough, when to route)
Research summarization: reading incoming AI/ML papers + news, writing 200-word weekly digests for Harnoor
Suggesting TITAN skill / hook upgrades based on observed friction
Monitoring for Claude Code / Anthropic product updates that TITAN should adopt

ECHO — red-team + QA (maintain 8-12h/week)

Was: Adversarial testing, 22-turn crisis script.

Now adds:

Every variant promotion from canary → production requires an ECHO red-team pass
Weekly automated crisis-script regression on the current default
Accessibility audit (monthly) — WCAG 2.2 AA + colorblind simulation + screen-reader passthrough
Voice-input red-team (once voice ships) — prosodic crisis cues

SCOUT — research (maintain heavy; the engine of the PhD docs)

No expansion. SCOUT is already heavy-loaded (3 PhD studies + all the memo agents this sprint). Maintain pace. Already delivering.

FORGE — code (maintain heavy)

Architectural change: every new FORGE PR must pass the Modularity Prime Directive review. Hardcoded config → env. AWS-specific calls → behind adapter. Hardcoded variants → register in variants.py. HERALD approves before merge.

ORACLE — intel (maintain light; it's a burst-use agent)

No expansion. ORACLE is fine at 5-6h/week — it's news-driven.

VAULT — memory (maintain light; steady state)

No expansion. VAULT is a keeper, not a creator. Scales naturally with incoming feedback patterns.

---

3. Work redistribution — concrete reassignments

| Current queued item | Was | Now |

|---|---|---|

| Onboarding flow design | nobody | GUIDE |

| Engineer onboarding doc | FORGE | GUIDE |

| /learn public pages | HERALD | GUIDE |

| UI microcopy lexicon | FORGE | GUIDE |

| Video walkthrough scripts | SCOUT | GUIDE |

| /sage dashboard build | FORGE | SAGE (design) + FORGE (impl) |

| Per-variant analytics | SAGE (future) | SAGE (now) |

| Transparency report | HERALD | SAGE |

| Cost anomaly detection | FORGE | SAGE |

| Model-tiering analysis | DARWIN (future) | DARWIN (now) |

| Research digests weekly | ORACLE | DARWIN |

| TITAN skill upgrades | DARWIN (sporadic) | DARWIN (weekly) |

| Claude/Anthropic product tracking | nobody | DARWIN |

| Variant-promotion red-team | nobody | ECHO |

| Weekly crisis regression | ECHO (ad-hoc) | ECHO (scheduled) |

| Accessibility audit monthly | nobody | ECHO |

| Architecture drift check | nobody | HERALD (Friday) |

| Variant status review | nobody | HERALD (Monday) |

| Interface approval gate for new adapters | nobody | HERALD |

---

4. Modularity Prime Directive — how each agent applies it

HERALD: Reviews every architectural proposal against the directive. Owns the Variant Registry + Admin Dashboard spec.

FORGE: Writes no code that violates it. Uses adapter interfaces everywhere. Pulls any hardcoded string into env or variants.py.

SCOUT: Researches alternative providers (LLMs, STT/TTS, DB, CDN) to validate adapter interfaces are future-proof.

GUIDE: Every UI string is externalized (variant + i18n ready) before the feature ships.

SAGE: Analytics tracks every variant combination. Cannot measure what isn't tagged.

DARWIN: Proposes new variants based on signal. Owns the promotion cadence.

ECHO: Every new variant at canary+ stage gets an ECHO pass before production.

ORACLE: Tracks competitor architecture choices — when we should adapt.

VAULT: Archives every prime directive + every variant history.

---

5. Research memos → architecture application

Every PhD memo written so far feeds directly into the new architecture. Cross-reference:

| Memo | Applies to |

|---|---|

| FEATURE-READINESS-STANDARD | Variant stage gates (experimental→production) |

| EMERGENT-CONSTELLATION-PLAN | A product surface — theme-tagging per user, per variant |

| SOUND-SCIENCE-MODES-RETENTION | Audio adapter interface; mode variant category |

| USER-FEEDBACK-SYSTEM | Rating variants (40-strong registry) + Chat Sentinel |

| ANALYTICS-DATA-COLLECTION | Bronze/Silver/Gold tiers + EMF schema + SAGE dashboard |

| MOBILE-APP-AUDIO-STRATEGY | Client-side on-device STT/TTS as yet another variant |

| GROWING-TREE-FEATURE | UI variant category (on/off, stage-model) |

| SILENT-INFINITY-AUDIT-DOCUMENT | Master spec — now binding |

| DATA-TRAINING-ANONYMIZATION | Data pipeline adapter layer |

| PERCEIVED-LATENCY-ANIMATION | UI variant category (ghost char, reveal style, etc.) |

| UI-LATENCY-MASKING-TRICKS | Same — tactical catalog for variant experiments |

| BUSINESS-MODEL-PRICING-v2 | Pricing variant category (tier structures A/B testable) |

| INCIDENT-RESPONSE-PLAYBOOK | Ops governance, not architecture but binding |

| RISK-REGISTER | Drives what adapters to prioritize (failover risk = multi-provider LLM first) |

| ENGINEERING-RUNBOOK | Operating procedure for architect-handoff |

| MODULARITY-PORTABILITY-VARIANT-ARCHITECTURE (in flight) | The master architecture memo |

| TOPIC-SUBTOPIC-EMOJI-PATENT-ANALYSIS (in flight) | IP strategy — influences brand moat |

| COMPETITIVE-MARKETING-STRATEGY | Goes to GUIDE for /learn page content |

| COMPETITOR-FUNNELS-FEATURES-LATENCY | Benchmarks for ECHO red-team |

| COMPETITIVE / BRAND / INVESTOR / INCIDENT | Cross-agent governance |

Every memo becomes an input to the new HERALD-architect role. HERALD is responsible for keeping them cross-referenced + acted upon.

---

6. Cadence — weekly rhythm with the expanded agents

| Day | Agent(s) | Deliverable |

|---|---|---|

| Mon | HERALD | variant status review — what's at each stage |

| Mon | DARWIN | weekly AI/ML research digest (200 words) |

| Tue | ECHO | red-team on any canary variant + weekly crisis regression |

| Wed | SAGE | cohort retention + cost + per-variant analytics report |

| Thu | GUIDE | one onboarding/learn/microcopy artifact |

| Thu | ORACLE | competitive intel brief (if anything material) |

| Fri | HERALD | architecture drift audit + release-note digest |

| Sat | SCOUT | deep-dive memo cycle (1/wk — topic rotates) |

| Sun | HERALD | leaderboard of the week |

---

7. Metrics for this update

Target by end of next sprint:

Each agent at <20h variance from the new target (no one at 1h, no one at 80h)
100% of new code committed passes Modularity Prime Directive review
At least 3 variants active (experimental / staged / canary) with telemetry flowing
/admin/variants dashboard MVP live
/sage private dashboard MVP live (even if sparse)
GUIDE ships first /learn page (pick: "how our crisis detection works")

---

— HERALD · architect + systems designer · 2026-04-21