How I help teams build with AI agents

Most teams I talk to can get good demos out of Claude but struggle to ship production software with it. The missing piece is a system around the agents: the context they load, documentation that stays accurate, tests that verify real behavior, and a workflow that lets several efforts run in parallel without colliding.

I’ve run engineering, product, and operations teams, and built software for every function of a startup, in the pre-AI era. Since early 2025 I’ve been figuring out how to do almost all of those things with AI at the core.

• Faxterra is a HIPAA-adjacent SaaS I build and operate with agents doing most of the work.
• At Fitwell, two non-technical founders run their business on operational systems they stood up themselves on a platform I built.
• At Wonderschool, the whole company works from one shared knowledge base and ops system through Claude.

This page describes the kinds of work behind those projects.

If you want to build this yourself, my AI coding best practices covers the core techniques. Working with me gets you reference implementations proven in production, plus help deciding what your situation needs.

What follows is a menu, roughly in the order teams tend to need things. Everything I deliver lands in one of two buckets: assets your team keeps and reuses (documents, skills, harness code, dashboards) and education (working sessions and recorded pairing on your real work).

Product & architecture evaluation

The goals:

• Claude knows the mission and vision of the business, and where to look to answer questions about the business, product, customers, and technical infrastructure.
• An architecture that meets the scalability, security, and data privacy needs of the business for the next year or two of expected growth.

Delivered as documents, developed through working sessions with your team:

A product and architecture evaluation. I review what you’ve built and what you’re building on: vendors, platforms, integrations. I check what each vendor actually covers against what you’re assuming it covers, and whether the composition holds up for what you’re trying to launch.

A hard-to-reverse decision record. Most architecture decisions are cheap to change later. A handful aren’t: build vs. rent, identity and source-of-truth models, where sensitive data lives (which drives compliance scope and cost), native app vs. web, hosting platform, and a data model that won’t need to be redone for next year’s plans. We decide these explicitly, with written rationale, before code depends on them.

Mission and context encoding. Discovery sessions with your team, encoded into structured documents in your repository: mission, product intent, business model, and the product knowledge that isn’t written down anywhere. Agents perform better with this context always loaded.

Knowledge management & tool access

The goal: Claude can do almost any task and answer almost any question that any employee has.

Two parts, both assets the whole team shares:

A unified knowledge base. Catalog every source of information Claude could benefit from when answering questions or building software: docs, wikis, tickets, dashboards, spreadsheets, the code itself. Move what makes sense into a unified knowledge base, and connect the rest through a consistent set of connectors, with guidance on where to go for what under what conditions. This is the heart of what we built at Wonderschool: engineers, PMs, designers, salespeople, and the data team all working from the same context through Claude.

Authentication into everything. Set up access to every third-party service Claude could benefit from: hosting, DevOps, monitoring, error tracking, product analytics, support, billing. Done consistently, so everyone on the team has the same access and the same capabilities instead of a patchwork of personal API keys.

The agent development harness

The goal: Claude can build, test, and ship a feature from nothing but a user story description.

I help build the initial AGENTS.md, skills, and supporting documentation that guide the development lifecycle, adapted from reference implementations I run in production. The pieces:

The instruction and context layer. A minimal always-on instruction set, a routing table that tells agents which docs to read before touching which code, and structured specs that hold the product and business knowledge that doesn’t live in code.

Documentation that can’t drift. Coverage tests that fail CI when a route, component, or user-facing fact is undocumented. Docs update in the same PR as the code, and agents are instructed to fix wrong documentation on sight.

Spec-driven building. Work plans with explicit phases and tests, picked up by agents that build them end-to-end in isolated worktrees, each with its own database branch and ports so multiple efforts run in parallel.

Verification. A Playwright harness engineered for agent reliability: direct auth session minting, dedicated test ports, seed and cleanup helpers. Agents also drive the real browser to verify what they just built with screenshots, click-throughs, and console output. A fast tier of checks runs on every change and a heavier tier on anything structural.

Workflow and ticket management. A system that de-conflicts human and agent contributors and supports agentic product planning. Depending on your situation this is an AI integration with a tool you already use (Linear, etc.) or a purpose-built board with guards that keep unattended agents from taking sensitive actions without human approval.

Claude Code coaching. Hands-on coaching in Claude Code best practices, working through your real backlog: model selection by task type, use of workflows, branching, session branching, rewind, and session context management. My AI coding best practices covers the foundations. This is the education half of the harness: the assets above only pay off once the team is fluent in them.

Operational automation

The goal: every meaningful event from users interacting with your system, positive or negative, bubbles up through an agent layer for decisioning and reaches humans in near real time.

Operational automation is mostly writing deterministic software rapidly with Claude: dashboards, admin tooling, scheduled jobs, and reconciliation. On top of that sit Claude skills for the repeatable tasks that genuinely need LLM judgment, like support triage and anomaly investigation. I supply reference implementations and review; the builds themselves become your team’s assets.

Typical builds: support ticket triage that enriches inbound issues and stages one-tap resolutions for the clear-cut cases; issue and error detection that turns anomalies into enriched tickets before customers report them; scheduled health checks across your infrastructure (deploys, database, SSL, error rates, vendor balances); security audits on dependencies, headers, and route auth coverage; reconciliation jobs that self-heal state drift; and a remediation playbook that requires human approval for anything outward-facing. Everything is inventoried in specs so nothing runs undocumented.

Where it fits, this culminates in an operator’s admin portal: real-time system health, revenue events, and a granular P&L in one homegrown view, instead of a patchwork of third-party dashboards and SaaS subscriptions. I run Faxterra exactly this way.

AI analytics & growth

The goal: rapid research and growth strategization, with an agent as your research and analysis pair.

Starter templates for analytics skills: reconstructing what users did from your analytics stream, funnel analysis, launch measurement. Plus a measurement plan covering what to instrument and which questions the data should answer. Most of the ongoing work here is pairing: sitting with an agent and your data to research a question, test a growth idea, or size an opportunity.

If this sounds like what your team needs: hello@gregmccullough.io