AI Agent Platform
Internal • April 2026Assisted
Sonnet
Haiku
pgvector
ellegentic from the model dropdown, or message Atlas in Slack.Copywriter → SEO Engineer → Formatter for social variants.intelligence or ask Atlas: "Generate a battle card for Seismic."store_finding for cross-agent access.prospector or ask: "Find 10 companies matching our ICP."enrich_lead, find_contacts, qualify_lead for deep-dive.closer or ask: "Prep me for a call with Acme Corp."herald directly.oracle or ask: "What's our enterprise pricing?"5sMax delay300sMultiplier2.0x exponentialLease900s (15 min)Heartbeat60sPoll2spublication_pipelinePost-approval prepcontent_generationSub-agent dispatchcompetitive_reportIntelligence reportslead_enrichmentProspector enrichment| Content Type | Auto-Approve | Min Score | Approvers |
|---|---|---|---|
| seo_technical | Yes | 80 | — |
| social_media | Yes | 85 | — |
| blog_article | No | 80 | elle |
| email_campaign | No | 80 | elle |
| comparison_page | No | 80 | elle + ian |
| ceo_content | No | 80 | elle + mel |
| video_content | No | 80 | elle |
| Table | Function | Key Fields | Index |
|---|---|---|---|
content_chunks | Content Brain | source_document, chunk_text, embedding(1024d) | ivfflat (20) |
semantic_memory | Agent vector memory | agent_name, content, embedding(1024d) | ivfflat (20) |
agent_memory | KV store + TTL | agent_name, key, value(JSONB), expires_at | idx_lookup |
agent_events | Event bus | source_agent, target_agent, event_type, payload | — |
agent_jobs | Job queue | job_type, agent_name, status, priority | idx_pending |
approval_queue | Review queue | content_type, brand_score, seo_score | idx_status |
feedback_log | Corrections | original_output, correction, embedding | — |
feedback_patterns | Patterns | pattern_type, corrected_pattern, confidence | ivfflat (10) |
agent_actions | Audit trail | action_type, tokens_used, duration_ms | — |
user_profiles | User data | user_id, role, team, preferences(JSONB) | — |
scheduled_tasks | Cron tasks | schedule(cron), next_run, enabled | — |
release_signoffs | Beta promotion | signoff_status, approved_by | — |
ANTHROPIC_API_KEYAWS_ACCESS_KEY_IDOLLAMA_BASE_URL| Tier | Model | Usage | Agents |
|---|---|---|---|
| Tier 1 | Claude Opus | Deep reasoning (rare) | Reserved |
| Tier 2 | Claude Sonnet | Content generation | ELLEGENTIC, Intelligence, Prospector, Herald, Oracle, Closer |
| Tier 3 | Claude Haiku | Fast routing | Atlas, Narrator |
| Local | Qwen3 | Free dev | Fallback |
emit("content_ready") → publication_pipeline → Job queued → HERALD distributes
| Framework | Why Rejected |
|---|---|
| LangChain | Generic abstraction, massive overhead, no content pipeline or brand checking |
| CrewAI | Too opinionated, no governance/approval workflow, no brand enforcement |
| OpenClaw | Critical security risk. 512 vulnerabilities identified (8 critical). 135K+ instances exposed to the internet. ClawJacked attack enables full remote takeover via a single link. 1-in-5 ClawHub plugins contain malware. Anthropic banned flat-rate subscriptions (Apr 2026). Meta banned it from corporate machines. A single instance can consume $1K–$5K/day in API costs. Not suitable for enterprise with company data. |
| NemoClaw (NVIDIA) | Alpha software — not production-ready. NVIDIA’s security layer on top of OpenClaw, announced at GTC March 2026. Improves containment (sandbox) but does NOT address behavioral governance — can’t verify if agent actions are correct or aligned with business goals. 391 open issues on GitHub. APIs subject to change without notice. Optimized for NVIDIA hardware (vendor lock-in). No content pipeline, no brand checking, no approval workflows. |
| Hermes | Not mature for production. Limited multi-agent support — no sub-agent hierarchy. No content pipeline or brand enforcement. Small community, limited documentation. Not designed for the kind of specialized agent orchestration (routing + delegation + governance) that our use case requires. |
| Temporal | Complex self-hosting (Cassandra + Elasticsearch), $200+/month, overkill. Good for durable execution but doesn’t solve content generation, brand voice, or multi-agent routing. |
| Inngest | Good for event-driven workflows, but no content pipeline, no brand voice, no multi-agent hierarchy or governance layer. |
| n8n (as brain) | We use n8n for GTM workflow orchestration (Clay, Apollo, Instantly), but agents do the reasoning. n8n = arms, agents = brain. Clean separation of concerns. |
| Optimization | How It Works | Savings |
|---|---|---|
| Tiered Model Selection | Haiku ($0.25/MTok) for routing, Sonnet ($3/MTok) for content, Opus only for complex reasoning. Most calls use the cheapest model. | ~70% |
| Single-Router Architecture | Atlas (Haiku) routes to 1 specialist agent. No multi-agent chain. LangChain/CrewAI create 3–5 LLM calls per request — we do 2. | ~60% |
| Content Brain Pre-Fill | 75K chunks of Spekit data injected as context. Agents don't need to "learn" from scratch — they already know the product, competitors, and customers. | ~40% |
| Brand Check = No Regen | Programmatic brand scoring (regex, not LLM). Catches AI-stink without burning tokens. Only regenerates if score < 80. | ~30% |
| Refinement Skip Research | When user says "make it shorter", we reuse previous research context. No web search, no Content Brain re-query. | ~50% |
| Tool Loop Cap | Max 10 iterations, 40-message cap, 3-error abort. Prevents runaway token consumption from stuck agents. | Safety |
| Sub-Agent Delegation | ELLEGENTIC dispatches to specialized sub-agents with focused prompts. Each sub-agent has a narrow scope → shorter prompts → fewer tokens. | ~25% |
