What Is It
Enceladus is a production platform designed and built from scratch which solves a problem the AI industry is actively wrestling with: how do you get multiple AI agents from different providers to coordinate real work on shared projects without collisions, data loss, or governance gaps? The platform manages 6 active production systems, has closed 831+ tasks across 48 features, spans 124,804 lines of code across 430 source files, and runs on AWS serverless infrastructure at ~$35/month. I am the sole architect, developer, and operator.
The core innovation is treating project management primitives — Features, Tasks, Issues — not as database records, but as ontologically defined objects with governed lifecycles, required evidence gates, and deterministic completion contracts. Every mutation flows through a single MCP server enforcing governance authorization on every write. Grounded in Clark & Chalmers’ Extended Mind thesis, the platform is constitutively part of the operator’s cognitive process — not merely infrastructure agents use — and no task advances on trust alone. The v4 architecture decomposes the monolithic tracker mutation layer into four bounded services along trust lines: Record Service, Lifecycle Service, Scoring Service, and Read API.
Architecture & Design Decisions
31 Lambda functions (Python 3.11, x86_64), 8 DynamoDB tables, 4 SQS FIFO queues, Neo4j AuraDB Free (graph-indexed search), CloudFront CDN, API Gateway HTTP v2, Cognito auth with Lambda@Edge
React 19 PWA with TypeScript, Vite, Tailwind CSS, TanStack React Query — mobile-first governance cockpit across all active projects with full primitive surface
Code-mode MCP: 4 governed meta-tools (search, execute, get_compact_context, coordination) over Streamable HTTP + OAuth 2.1/PKCE — 89% schema reduction vs. 50-tool raw mode, which is preserved for local Claude Code desktop sessions
21 GitHub Actions workflows, nightly SHA-256 parity audits, secrets guardrail, 13 deployment types with semver changelog; component registry enforces deployment transition arcs at the infrastructure level
Coordination API with dispatch heuristics and graph-indexed tracker search across Claude, OpenAI Codex, and AWS Bedrock; 10 typed relationship edges with weight, confidence, and provenance
Systems Thinking & Key Innovations
Institutional Memory
The knowledge graph implements a framework for fast access (500ms) to Applied Knowledge (Working Memory) while the full Acquired Corpus (Long Term Memory) remains available for graph analysis and exploration. Applied Knowledge is a mathematically computated subgraph query result serving token optimized context to each agent session for the present intent. Three-signal hybrid retrieval (vector similarity + graph traversal + keyword RRF) achieves ~87% accuracy versus ~65% for keyword-only. A Memory Consolidation Lambda extracts recurring patterns from agent-to-agent communication nightly and proposes Lesson candidates for human approval based on common patterns.
Exclusive Checkout Service
Prevents agent collisions via atomic task ownership — only the owning session can advance status. Child tasks support parallel dispatch. Commit Approval IDs (CAI) gate code completion; Commit Complete IDs (CCI) are generate by Enceladus and validated in PR bodies by GitHub Actions before merge. A checkout-service assistant subsystem auto-remediates misclassified tasks by relaxing deployment arcs when evidence confirms the mechanism, and can never tighten them.
Evidence-Gated Lifecycle
Task state machine (open → in-progress → coding-complete → committed → pr → merged-main → deploy-success → closed) requires proof at every gate: commit SHAs validated against GitHub API, PR merge timestamps within 60-second tolerance, deployment evidence from GitHub Actions Jobs API with 7 validated fields. The component registry enforces the most-restrictive deployment arc across all components a task touches — infrastructure-level governance agents cannot bypass.
Governance as Architecture
SHA-256 governance hash required on every write mutation. MCP-API boundary policy ensures no tool handler directly accesses DynamoDB business tables, preventing transport-specific behavior drift. Write-source attribution on all mutations enables the audit Lambda to detect anomalies and alert via SNS. A CI-gated governance data dictionary enforces authoritative field semantics at both runtime and deploy time.
Token Economy Design
Every design decision weighs token cost. The code-mode MCP interface delivers an 89% tool-schema reduction. Graph-indexed tracker search collapses 20–50 sequential record lookups into 1–2 Neo4j traversal queries. Context nodes wrap every tracker record in mathematically-scored metadata enabling greedy knapsack context packing within token budgets. Prompt caching (90% discount), strategic model selection, and batch API routing (50% discount) compress operational AI cost by an estimated 60–80% overall.
Event-Driven Pipelines
DynamoDB Streams → EventBridge Pipes → SQS FIFO (natural debounce via 5-min visibility timeout) → Lambda feed publisher → S3/CloudFront invalidation. The same stream architecture drives the Neo4j graph sync with graceful degradation — DynamoDB mutations are never blocked by graph failures, and agents receive structured fallback hints when the index is unreachable.
By The Numbers
Bottom Line
Enceladus demonstrates that a single operator with the right abstractions and governed AI agents can build and maintain production systems at a scale and quality level that would traditionally require a team. The platform’s patterns — ontological entity discipline, evidence-gated state machines, exclusive checkout ownership, component-enforced deployment arcs, and mathematically grounded institutional memory — are architecturally ahead of where the multi-agent framework ecosystem (LangGraph, CrewAI, AutoGen) currently delivers, and represent the governance layer the industry is independently converging toward.