Build Log
Built in Phases. Every One Shipped.
A chronological record of everything we've built, from the first auth endpoint to governed execution across legacy systems. Every phase in production.
v2.5: July 2026
Governed Execution, Proven in Production
- The governed execution engine ran end-to-end in production: an agent action flows through a policy check, human approval, a blind-relay vault that grants access without ever exposing the secret, execution, and a sealed audit record
- First human-in-the-loop approval fired in production: an agent action stopped at the gate, waited, was approved by a person, then executed and sealed to the record
- Approval notifications. The moment an action needs a human decision, the person is paged with a direct link, so nothing waits unseen
- Federal governance primitives: NIST 800-53 control mapping, OSCAL audit export, and an air-gap deployment edition for on-premise and government environments
- Hardening: full dependency-tree audit clean, security-sensitive database functions locked down, and the immutable-journal and job-queue controls verified under adversarial tests
v2.4 · April–May 2026
Agent Runtime Primitives
- Intent-category routing: agents and workflow steps declare a category; per-org category→model mapping replaces hardcoded model strings, takes effect with no redeploy
- Step-scoped integration activation: workflow steps bind integrations; runtime enforcement prevents agents from calling unbound integrations; per-step cost attribution via execution_id + step_node_id
- Loop-until-done step type: first-class long-running agent loops with three hard guards (max-iteration ceiling, monthly budget cap, idle detector) and journaled abort reasons
- Hierarchical memory: cross-agent memory resolves along scope hierarchy (org → estate → agent → workflow → step) with auto-injection on step entry, token-budget capped, RLS-isolated
- Hash-anchored state mutations: version_hash on vault/memory/integrations; stale writes return HTTP 409 instead of silently clobbering; SDK 0.5.0 exposes readWithHash / writeExpecting helpers
v2.3: April 2026
Autonomous Pulse System
- 5-minute Pulse cron with situation-hash gating: 90% LLM call reduction when platform state is unchanged
- 11 agent crons wired with heuristic observation recording: agents notice anomalies between scheduled runs
- Append-only agent journal with monthly range partitioning: immutable audit trail for governed autonomy
- 4-phase Reverie nightly consolidation: merges, deduplicates, and prunes cross-agent memory overnight
- Confidence-gated Telegram notifications: immediate alerts for critical events, daily digest for everything else
Public Beta · March 2026
Oceum Enters Public Beta
- Every phase shipped to production, from first auth endpoint to governed execution across legacy systems
- 75 live serverless functions, 62 pages, 927 tests passing across 88 files, 26 database tables
- 10 autonomous agents: Briefing, Content, Deploys, Ledger, Pipeline, Revenue, Security, Triage, Uptime, and Orion (AI COO with RAG knowledge layer)
- v2.0 Execution Engine: 5 protocol adapters (REST, SOAP, SFTP, Webhook, JDBC) with blind-relay credential injection
- Full governance layer: graduated autonomy, approval gates, audit trails, and replay of failed executions
- Enterprise self-hosting via Docker with Postgres adapter and source sync pipeline
- iOS app approved on App Store, Drift Engine for autonomous content, Recall for semantic memory compression
- SOC 2 compliance documentation, Trusted Autonomy® trademark filed (registered Aug 2026, Reg. 8385833)
Phase 48 · March 2026
Execution Dashboard
- Live execution queue with status indicators (pending, running, success, failed) with real-time updates
- Inline approval queue: operators approve or reject governed actions with full connection context
- Connection health dashboard with adapter status and last-successful-execution timestamps
- Execution history with filtering by connection, adapter type, status, and date range
Phase 47 · March 2026
Governance Bridges
- Workflow composer steps invoke governed connections with approval gates. A workflow step calling SAP goes through the execution pipeline
- 28 existing OAuth integrations can optionally route through governance for audit trails and approval
- Policy engine evaluates governed execution requests from both workflow and integration sources
Phase 46 · March 2026
JDBC Adapter + Data Mapping
- Database adapter: parameterized queries against Postgres with connection pooling (extensible to MySQL, MSSQL, Oracle)
- Read-only by default: write operations require explicit governance approval
- Canonical business objects (customer, order, invoice) with field-level mappings between legacy formats
- JSON Schema validation and schema drift detection when source systems change unexpectedly
- Self-hosted only: adapter refuses execution in Vercel serverless
Phase 45 · March 2026
Webhook Receiver + SFTP Adapter
- Webhook endpoint receives inbound events from legacy systems with HMAC-SHA256 signature verification
- Verified webhook events trigger agent workflows through the governance pipeline
- Dead-letter queue for undeliverable webhook events
- SFTP adapter: upload, download, list files with vault-injected SSH keys or passwords
- File change detection for poll-based workflows: self-hosted only
Phase 44 · March 2026
SOAP Adapter
- WSDL parsing and SOAP envelope construction with XML namespace support
- WS-Security headers for authentication against SAP, Oracle EBS, and government endpoints
- XXE and SSRF hardening: external entity resolution disabled, WSDL fetch restricted to approved hosts
- Same adapter interface contract as REST, no special execution engine handling
Phase 43 · March 2026
Execution Foundation
- Postgres-backed job queue with retry, exponential backoff, and dead-letter for permanently failed executions
- Blind-relay credential injection: vault secrets decrypted at execution time, injected into protocol-specific auth, cleared from memory
- REST adapter: all HTTP methods with configurable headers, body, query params, timeout, and circuit breaker
- Output sanitization: credentials never appear in logs, execution records, or error messages
- Works in all three deployment modes: Vercel cron, Docker setInterval, and air-gapped
Phase 42 · March 2026
Oceum Recall · Compressed Semantic Memory
- Compressed vector memory for agentic platforms: 7.7x storage reduction at 3-bit quantization with >0.99 cosine fidelity
- Adaptive bit-width selection: memories scored by recency, access frequency, and agent priority
- Per-org rotation key encryption via HKDF: tenant-isolated similarity search without decryption
- Drift-aware decay: tiered compression (hot→warm→cold) with access-based promotion back to full fidelity
Phases 40–41 · March 2026
Security Hardening + Compliance
- 28-fix security audit across all platform systems: org scoping, SQL injection, cascading deletes, token revocation
- Full test coverage: 329 tests across 22 test files (now 2,100+ tests)
- Legal compliance suite: Privacy Policy, Terms of Service, DPA, SLA, Security Policy, all published
Phase 39 · March 2026
Governance as a Service
- Policy engine: governance rules (budget caps, approval thresholds, content restrictions) across all agents
- Real-time policy evaluation: agents check actions against active policies before executing
- Approval callbacks with configurable escalation paths
- SDK integration: one-line policy evaluation from any agent or external system
Phase 38 · March 2026
Self-Improving Agents
- Agents automatically choose higher-performing prompt variants based on outcome data
- Weekly optimization crons: prompt refinement, strategy review, gap detection, cross-agent retrospective
- Self-improve dashboard with Prompt Lab, Strategy, Gaps, and Retrospective views
Phase 37 · March 2026
Reasoning Chains + Audit Trails
- Every autonomous decision logged with data considered, options weighed, action chosen, and outcome
- Confidence scoring and counterfactual analysis: what if the agent chose differently?
- Decision chaining: multi-step reasoning chains with full traversal views
- Enterprise compliance: complete audit trail for every autonomous action
Phase 36 · March 2026
Agent Templates
- Pre-configured agent bundles: Marketing, DevOps, Support, Full Stack, one-click deploy
- Custom templates: save your agent configuration for other organizations
- Compatibility checking and smart deduplication
Phase 35 · March 2026
Workflow Composer v2
- Workflow nodes dispatch to real agents, no more simulated returns
- Async execution tracking with per-node progress and live status updates
- Workflow versioning and scheduled triggers
Phase 34 · March 2026
Self-Healing Foundation
- Watchdog cron: checks agent heartbeats every 5 minutes, auto-restarts stale agents
- Credential expiry alerts and event bus health monitoring
- Retry and fallback patterns wired across all external integrations
Phase 33 · March 2026
Goal Decomposition Engine
- Set high-level goals in natural language: AI decomposes into actionable subtasks with agent assignments
- Recursive task trees with progress tracking, priority badges, and deadline countdowns
- Watchdog auto-evaluates stalled goals and approaching deadlines
Phase 32 · March 2026
Universal Graduated Autonomy
- Per-agent, per-action trust scores (0-100) for 6 action types: social, email, modify, delete, spend, API
- Trust-gated autonomy: low trust forces supervised mode regardless of configuration
- Auto-escalation on success streaks, auto-revocation on failures
- Governance gate: high-risk decisions require approval, medium-risk auto-audited
Phase 31 · March 2026
Agent Event Bus
- Event-driven coordination across all agents: emit, consume, claim with first-write-wins
- 15 event types wired across 11 cron agents
- Dashboard event activity feed for real-time cross-agent coordination visibility
Phase 30 · March 2026
Runtime Policy Engine + Granular RBAC
- Every execution evaluated against deterministic rules before touching production systems
- Policy types: hard limits, scope-based authorization, contextual triggers for human approval
- Behavioral drift detection: anomaly alerts, scope-boundary warnings, reasoning drift monitoring
- Granular RBAC: agent identities with bounded scopes, human roles (builder, operator, approver, auditor, admin)
Phase 29 · March 2026
SOC 2 + GDPR Controls
- MFA/2FA with TOTP and backup codes
- Token revocation, account lockout, password complexity enforcement
- Login audit trail and admin action audit logging
- GDPR data export (Art. 20) and log retention enforcement
- Row Level Security on all tables, CI/CD pipeline with automated security audits
Phase 28 · March 2026
Enterprise Governance API
- Governance API with 12 actions: connections, adapters, executions, approvals, audit trails
- 10 connection types: SFTP, JDBC, SOAP, REST, webhook, message queue, and more
- Execution modes: auto, approval_required, dry_run with state machine transitions
- Agents cannot approve their own actions: human-in-the-loop enforced
- Replay engine for retrying failed executions from checkpoint
Phase 27 · March 2026
Multi-Org Support
- Users belong to multiple organizations with per-org roles
- Instant org switching. Each org gets its own agent estate, vault, drift engine, and billing
- Per-org encryption keys for complete tenant isolation
Phase 26 · March 2026
Orion Training Infrastructure
- Every Orion chat logged as training data with user feedback signals
- Synthetic data generator: 28 categories, 400+ prompts, 15,000 training examples
- Alpaca-style JSONL export for fine-tuning pipelines
Phase 25 · March 2026
Security Hardening
- Full systems audit: 13 findings identified and fixed
- Prompt injection defense, API key rotation, fetch timeouts on all external calls
- Per-agent rate limiting and server-side memory TTL enforcement
Phase 24 · March 2026
Drift Engine Asset Library
- User-uploaded image library: brand photos, product shots, lifestyle images
- Intelligent asset selection with tag-based scoring and freshness balancing
- Asset preference toggle: Prefer My Assets / Prefer AI / My Assets Only / AI Only
Phase 23 · March 2026
Enterprise Self-Hosted Deployment
- Complete self-hosted deployment package with Docker Compose
- JWT license validation: Community, Team, Enterprise tiers
- Zero Supabase dependency runs on raw Postgres only
Phase 22 · March 2026
Database Adapter Pattern
- Dual-adapter architecture: Supabase for cloud, raw Postgres for enterprise
- Chainable SQL compiler with 12 filter types: zero consumer files changed
- Lazy Supabase import. Enterprise builds don't need the Supabase SDK
Phase 21 · March 2026
API Hardening + Enterprise Testing
- Standardized response envelope across all API endpoints
- Rate limiting and input validation on every authenticated endpoint
- Sentry observability, OpenAPI 3.0 spec, and CI/CD pipeline
- SDK v0.2.0 with vault proxy, memory search, and workflow trigger
Phase 20 · March 2026
Voice Chat + Agent Estate Expansion
- Voice chat with speech-to-text and text-to-speech in Orion
- Revenue (MRR tracking) and Deploys (deploy monitoring) agents added to the agent estate
- Gmail API integration for outbound outreach
Phase 19 · March 2026
Blind-Relay Vault Proxy
- Blind relay: agents make API calls without seeing raw credentials
- Domain allowlist per vault token prevents credential exfiltration
- SSRF prevention: blocks localhost, private IPs, and metadata endpoints
Phase 18 · March 2026
AI Image Generation
- GPT Image-1 for Drift Engine content images
- Automatic image generation during content publishing
Phase 17 · March 2026
Social Publishing Pipeline
- LinkedIn, Instagram, and Facebook publishing pipeline
- OAuth connections for Meta and LinkedIn
- Drift content auto-publishes to connected social channels
Phase 16 · March 2026
Drift Engine
- Autonomous content research, generation, and publishing via 3 coordinated crons
- Reputation scoring system governing content quality
- Kanban UI for content pipeline management
Phase 15 · March 2026
Brand Context + Onboarding
- 5-step onboarding questionnaire for personalized agent behavior
- Brand-first context system with voice and tone settings per-org
Phase 14 · March 2026
Admin Portal
- Cross-org admin interface for Orgs, Users, Agents, and Logs
Phase 13 · February 2026
Beta Launch
- Password reset flow, pricing page, Slack + email integrations
- Published SDK to npm: agent estate flipped to fully autonomous
Phase 12 · February 2026
Pipeline Lead Pipeline
- Dual pipeline: user leads with ICP scoring, investor leads with founder tone
- HN Scout discovers relevant Show HN stories automatically
Phase 11 · February 2026
Autonomy + Agent Estate Activation
- 3-tier autonomy with whitelisted actions and readiness gates
- Notification system across Telegram, Slack, and email
- Orion upgraded to COO-level context awareness
Phase 10 · February 2026
Dogfooding
- All 10 agents deployed to manage Oceum's own operations
- Security auditing the platform that runs Security
Phase 9 · February 2026
SSO + SDK
- Google and Microsoft SSO via OpenID Connect
- Published SDK to npm: TypeScript, zero dependencies, multi-runtime
Phase 8 · February 2026
Supabase Migration
- Migrated data layer from Notion to Supabase Postgres
- RLS, cascading deletes, 10x faster queries
Phase 7 · February 2026
Visual Workflow Builder
- Drag-and-drop canvas with Bézier connections
- 5 node types, conditional logic, one-click execution
Phase 6 · February 2026
OAuth Integration Layer
- OAuth flows for Slack, Google, Microsoft, Notion
- Token auto-refresh and per-org key management
Phase 5 · February 2026
Orion AI Chat
- Context-aware copilot on every page: knows your agents, integrations, and logs
- Natural language interface for platform management
Phase 4 · January 2026
Agent Governance
- Agent registration, status tracking, pause/resume/delete
- Agent dashboard with real-time status and webhook API
Phase 3 · January 2026
Integration Connectors
- Token management with auto-refresh across 28 integrations
Phase 2 · January 2026
Onboarding + Dashboard
- Multi-step onboarding, agent stats dashboard, activity logging
Phase 1 · January 2026
Auth + Foundation
- JWT auth, org management, Vercel serverless architecture