What I use

Editor + agent loop

The two surfaces that touch every line of code I ship.

• Cursor

  Primary IDE.

• Claude Code v2.1.117+source

  Agent loop with sandbox subprocess isolation enabled (CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1, PID-namespace isolation). verdict (the rubric I ship) is verified against this version.

• tmux + neovim

  Long-running sessions, paired with Cursor for review-on-cursor and ship-on-tmux.

Models in production rotation

Cascading router — simple → small models, complex reasoning → Opus / Sonnet / GPT-5.5 Pro. Gemini 3 Pro (preview) is eval-only.

• Claude Opus 4.7

  Strategy + planning + nuanced long-context. Internal id: claude-opus-4-7.

• Claude Sonnet 4.7

  Default workhorse for the Agentify orchestrator.

• Claude Haiku 4.5

  Routing + pre-classification. Internal id: claude-haiku-4-5.

• GPT-5.5source

  OpenAI default. Internal id: gpt-5.5.

• GPT-5.5 Pro

  Hard-reasoning cases. Internal id: gpt-5.5-pro.

• Gemini 3 Pro (preview)source

  Eval-only — not in production rotation. Used to triangulate when Anthropic + OpenAI disagree.

• Mistral Medium 3.5source

  Active eval (GA 2026-05-01). Reasoning-toggleable; agent-runtime evals only — not yet in production routing fallback chain.

Runtime

• Python 3.13

  FastAPI, Pydantic v2, asyncio everywhere.

• Node 22 LTS

  Frontend + Edge runtimes.

• Next.js 16.x

  App Router, RSC, Edge + Node runtimes per route. Pinned ≥ 16.2.4 for CVE-2026-23869 + CVE-2026-29057.

• TypeScript strict

  All frontend + most scripts.

• Rust

  mnemo + critical infra paths — anywhere I need tight memory control + predictable latency.

Infrastructure

• Vercel

  Frontend + Edge functions for the portfolio + agent-airlock demo surfaces.

• AWS Lambda + ECS

  Long-running agent workloads at Attri.ai.

• AWS Bedrock Managed Agents (Codex / GPT-5.5, limited preview)source

  Limited preview, Apr 28 2026. Wraps any Bedrock-Codex / GPT-5.5 invocation with the agent-airlock sandbox + capability-lease envelope. Optional bedrock_invocation_arn field shipped in /api/quote/schema.json v0.2.

• E2B microVMs

  Code execution sandbox for unverified agent-written code. Zero production incidents to date.

• Azure WebPubSub

  Real-time multi-tenant collaboration in Agentify.

• PostgreSQL + pgvector

  Primary store. Vector search lives next to the source of truth.

Agent identity (3-layer)

The non-human-identity stack as of mid-2026. See /writing/capability-lease for the manifesto + working wire format.

• Okta for AI Agents (NHI binding, opt-in default-off, GA 2026-04-30)source

  Okta Universal Directory now hosts non-human identities (NHIs) for agents alongside humans. Lifecycle, audit, revocation are first-class. Opt-in module in the consulting starter from today; existing engagements get a one-line patch that binds capability leases to Okta NHI tokens.

• W3C DID + Verifiable Credentialsource

  Cryptographic identity layer. /api/quote already accepts agent_identity.did + verifiable_credential per draft-2020-12 schema; verification on the v0.4 roadmap.

• Capability-lease envelope (per-call authorization)source

  Time-bounded, scope-bounded lease per call with audit-emit channel. Anthropic Project Deal · Cloudflare Mesh · Cisco AWI converge on this shape.

Security tooling (open-source, mine)

I ship and dogfood the agent-security stack I write about. Both repos are MIT-licensed.

• agent-airlocksource

  Runtime firewall — ghost-arg stripping, PII masking, RBAC, E2B sandboxing, cost circuit-breakers.

• agent-audit-kitsource

  SAST scanner for agentic systems — OWASP Agentic Top-10 + MCP Top-10 + 11-framework compliance reporting (SARIF).

• verdictsource

  7-dimension quality rubric for Claude Code; auto-hooks into the agent loop.

Observability

• Datadog

  Tracing + metrics across the multi-agent fleet.

• Sentry

  Error capture, frontend + backend.

• PostHog

  Product analytics + feature flags.

• OpenTelemetry

  Trace surface for agent steps.

Hardware

• MacBook Pro M-series

  Primary daily driver.

• External 4K + ergonomic mech keyboard

  12+ hour days demand it.

Physical AI (VAJRA build)

Pinned versions for the build-in-public Physical AI track. Status: under construction; see /vajra for the live build log.

• Jetson Orin Nano Super (8 GB)

  Edge compute. JetPack 6.3 (L4T r36.4.x).

• Intel RealSense D435i

  Depth + IMU sensing — sourced, integration in progress.

• Custom rover chassis

  Differential-drive base, 5-DOF arm + parallel gripper. In fabrication.

• ROS 2 Jazzy 0.13source

  Migrating from Humble; Nav2 stack on top.

• NanoOWL @ 0.4.2source

  TensorRT-optimized OWL-ViT for open-vocabulary detection. Pinned commit-sha checked into the build manifest.

• GR00T N1.7 (reference impl)source

  NVIDIA April 2026 release — used as architectural reference, pinned to a specific commit-sha in the build manifest.

• Cosmos 3 sim assets v3.1

  Synthetic-data factory for affordance training.