AI Agent Control Plane

Conductor

A production-grade orchestration system that deploys AI agent swarms to turn your GitHub backlog into a stream of reviewable pull requests. Adversarial review, parallel portfolio execution, token optimization, and human gates at every critical decision point. Battle-tested across a 100+ issue production codebase.

Core Loop
Issue in. Pull request out.
Issue → PM Interview → Plan Gate → Human Approval → Agent Swarm → Adversarial Review → PR → Merge
Conductor doesn't replace developers. It eliminates the 80% of mechanical execution work so your team focuses on the 20% that requires human judgment: approving the plan and reviewing the result.
Swarm Architecture
Three specialized agent roles, fail-closed phase gates
The orchestrator deploys agents with structurally enforced authority boundaries. Agents cannot commit, push, transition issue states, or interact with users. Post-return verification catches violations. Every phase gate requires specific evidence before progression.

Orchestrator

The main conversation. Enforces process, manages user interaction, deploys agents, verifies outputs, controls state transitions.

Planner Agent

Explores the codebase and collaboratively builds implementation plans. Plans are adversarially reviewed before human approval gate.

Developer Agent

Implements the approved plan in an isolated worktree. Code is adversarially reviewed before it can proceed to PR.

Adversarial Review System
Dual-AI review with instance-verified findings
Not rubber-stamp "LGTM" reviews. Claude drafts, a second AI independently red-teams. Both must approve. Every finding requires proof that the problematic pattern actually exists in the codebase. No theoretical concerns allowed.
  • Review ledger system. Findings tracked across review rounds with statuses (justified, withdrawn, fixed, open). Reviewers don't re-raise resolved issues.
  • Seven review principles. Deep Verification, Scope Verification, Failure Mode Analysis, Comment Skepticism, Infra Parity, Test Depth, Scope Mixing.
  • Severity with evidence. BLOCKING findings require search-verified proof. No theorizing. Exception only for financial-risk contract code.
  • Structural enforcement. Analysis-only mode forbids write operations at the tool level, not just behaviorally.
Portfolio Orchestrator
Multiple issues, in parallel, across isolated environments
Each issue gets its own tmux window with an isolated Claude session, Git worktree, and automatically offset ports. A real-time status bar shows which sessions need human input. Desktop notifications fire when an agent is blocked.

Worktree Isolation

Every issue runs in its own Git worktree with calculated port offsets. Multiple development environments run simultaneously without conflicts.

Hook-Driven Status

Pre/post tool hooks track session states in real time: needs-input, needs-permission, running, idle, complete, crashed.

Shared MCP Infrastructure

Instead of each session spawning its own MCP servers, a shared proxy layer serves all sessions. Health-aware fallback to per-session stdio if a server fails.

Discovered Work Handling

When agents find out-of-scope work, they stop immediately and return structured payloads. The orchestrator creates separate issues. No scope creep.

Token Optimization
81-89% token savings through intelligent compression
Analyzed 1,550+ real commands to identify compression opportunities. Build outputs, test results, and CI logs are intelligently compressed before reaching the AI context window. Measured data, not estimates. Per-request cost tracking across all agent pools.
Security
Agents are treated as untrusted input
  • Pre-tool command guards. Hook system validates every command before execution. Secret detection on both inputs and outputs.
  • Deny-by-default network egress. Sandbox blocks all outbound except an explicit allowlist.
  • All GitHub writes audited. Every write logged with target, payload hash, and policy decision via an outbox pattern.
  • Emergency stop. Pause any run, bulk cancel a project, or disable all agent execution system-wide.
  • Redaction. API keys, tokens, and passwords automatically stripped from logs, comments, and stored artifacts.
Production Stats
100+
Issues processed
126+
Passing tests
81-89%
Token savings
6
Custom AI skills
Also Built On This Platform
AI-powered workflows beyond development

Automated Weekly Reports

GitHub Actions generate structured JSON from project data. AI produces executive narratives with week-over-week deltas, risk assessments, and stakeholder framing. Auto-filed as PRs.

Smart Contract Upgrades

Orchestrated upgrade pipeline: pause, deploy, verify, update registry, resume. Full audit trail with rollback support. No manual steps.

Tech Stack
Built for production, not demos
TypeScript (strict mode), Next.js 14+, SQLite (WAL mode), Redis + BullMQ, GitHub App integration, Docker Compose. Monorepo with shared, web, and worker packages. Claude Code skills system with custom hooks, MCP servers, and tmux portfolio orchestration.
Discuss a build like this