Orchestration, Triage Gateway, and Guardrails
Multi-Agent Orchestration App (Temporal + LangGraph)
- Up to 1,000 parallel agents: Temporal durable workflow engine (distributed control plane) + LangGraph graph-based state runtime (local agent loop).
- Git worktrees dynamically provisioned per agent thread — eliminates concurrent file-modification conflicts, keeps changes separated.
- Plan Mode discipline: high-capability model in a planning container analyzes service dependencies, compiles explicit execution playbook (
plan.md) — ordering constraints, transactional boundaries, explicit human-escalation conditions. After plan validation, code modification delegated to smaller/faster/cheaper executor models. - Prevents sequencing errors: e.g. writing DB columns before updating insertion logic; treating shared DB tables as independently owned by an extracted microservice.
Two-layer structure:
LangGraph local loop: Analyze Code → Write Code → Run Compiler →(error? loop)→ Execute Tests
↓ durable step transition
Temporal control plane: Activity 1 (local agent run) → Activity 2 (triage gate) → Activity 3 (system rollout)
transactional durability · replay memory ledger · fault isolation
Intelligent Question Triage & Routing App
At 1,000 agents, clarifying-question volume overwhelms teams. Gateway between fleet and humans:
- Intercept + standardize every agent query into structured JSON: target microservice, file paths, proposed change, calculated confidence score.
- Semantic deduplication: vector embeddings of question text, real-time clustering. New question matches active cluster → auto-apply previous human decision (resolves similar blocks across microservices).
- Dynamic risk scoring for unique questions: low-risk (score < 0.70) → automated checkpointer gates; high-risk (≥ 0.70: DB migrations, authz changes) → human stakeholders.
- Ownership routing: map microservice → owner via enterprise directory; deliver concise triage alert via Slack/Jira — side-by-side AST-based diff, trade-off summary, risk score, one-click approve/edit/reject.
Measured effect: human interruptions reduced from 8 per task to 3 checkpoints.
Autonomous Quality & Security Guardrail App
Automated gate inside agent execution loop + CI/CD:
- Code Health score (1–10) via MCP server to a code-health platform (CodeScene-style): structural complexity, nesting depth, class size. Agent must iterate in local validation loop until file reaches ≥9.5 “AI-ready” before commit.
- Per-PR scans:
- Semgrep + OSV-Scanner — security vulnerabilities, outdated dependencies
- jscpd — copy-paste duplication
- Anti-pattern checkers — TypeScript
anyabuse, “comment floods” (excessive model-added comments), “ghost files” (redundant files from incomplete context)
- Agent-generated PRs still route through human reviewers for final approval: unguided code-review-agent comments are noisy (60.2% fall in a 0–30% signal range) and raise reviewer cognitive load — combine automation with human architectural judgment.