Skip to content

Engineering Process

Every line of code verified by two independent AI models. Every change passes through specialized agents. Every build validated by IDE-native static analysis.


Autonomous Multi-Agent Engineering

MCP-Orchestrated Development Workflow — click any stage to explore

Research
Plan
Implement
Review
Quality Loop
Ship
Cycle 1continuous loop

Codebase Research & Exploration

Phase 1 — Map before you move

Every task begins with structured research. Up to 3 parallel sub-agents explore the codebase simultaneously — locating files, analyzing signatures, and mapping usage patterns through the IDE integration layer.

Tools
Fast sub-agents (exploration)IDE integration (symbol search, inspections)Static analysis layer
1.Define 3-5 research questions
2.Launch parallel sub-agents
3.Map dependencies via MCP
4.Synthesize to thoughts/research/
5.Cross-model validation (optional)
Output

thoughts/research/YYYY-MM-DD_<topic>.md — structured findings with file:line references, architecture context, data flow, and gotchas

Impact Multiplier: Research review = 100-1000x impact. Catching wrong assumptions here prevents cascading errors downstream.

Dual-Model Verification·IDE Integration·Multi-Agent Fleet·Custom Skills

Single-model AI development has three fundamental limitations: single point of failure (one model’s blind spots propagate), context window exhaustion (large codebases exceed session capacity), and no structural guarantees (conventions, tests, and edge cases get skipped). This architecture addresses each:

ProblemSolution
Single model blind spotsDual-model verification — two frontier models cross-verify every change
Context exhaustionFile-based context management — research writes to persistent files, implementation reads only what’s needed
No structural guaranteesAgent constitution pattern — specialized agents with exclusive file ownership and non-negotiable rules

Every task runs through a structured pipeline with verification at each stage:

  • Research → Planning → Implementation with review gates between each phase
  • Impact multipliers: Research review catches wrong assumptions early (highest leverage). Plan review validates approach. Code review catches bugs.
  • Context discipline: Compact after every 2-3 implementation phases, use sub-agents for exploratory searches

Every task runs both models in parallel. Neither trusts the other.

TaskPrimaryVerifierMode
Architecture & root cause analysisModel AModel BParallel, compare
Security audit, consensus reviewModel AAdversarial challengeParallel, compare
Implementation & bug fixesModel ACross-verificationParallel, cross-verify
Test writingRegression + coverageEdge-case + adversarialParallel, merge
Code review & pre-PR gateStandard reviewAdversarial reviewParallel, merge findings

AI agents connect directly to the IDE runtime via the Model Context Protocol. Every build, inspection, and refactoring goes through the IDE — not CLI approximations.

Capabilities include: full compilation with error reporting, static inspections, semantic symbol lookup, IDE-aware rename refactoring, test execution, project-wide search, and code formatting — all through the live IDE session.


Every commit must pass all gates. No exceptions, no overrides:

  • Compilation clean (IDE build, not just CLI)
  • Static analysis clean for all modified files
  • All tests passing
  • No critical or high-priority items introduced
  • No forbidden patterns (thread-safety violations, raw types, catch-all exceptions)
  • Null-safety annotations on all new public API
  • Cross-model review passed

Every layer of the build pipeline catches different categories of defects:

LayerWhat It Catches
Compile-time analysisType errors, API misuse, concurrency bugs, null-safety violations
Bytecode analysisConcurrency bugs, resource leaks, security anti-patterns
Dependency scanningCVE detection on every build
Code coverageLine, branch, instruction metrics
IDE-grade inspectionsSame rules in CI/CD as local development
Container buildsMulti-stage builds, health checks, metrics exposure

Real metrics from production codebase — not estimates, not projections:

MetricValue
Production code105,000+ lines of Java 25
Source files447 across 5 Maven modules
Test coverage637 test files, 14,100+ individual tests
Protocol messages108 binary message types
Data structures14 distributed + 5 CP primitives + 4 CRDTs
Static analyzers6 independent analysis layers

Tech stack: Java 25 (preview features), Spring Boot 4.0.5, Raft Consensus, mTLS/RBAC, Virtual Threads, custom binary protocol, Docker multi-stage builds, Prometheus metrics.


Browse the LoomCache documentation to see the product built by this process, or the Agentic Development page for the service overview.