Systems operationalAgent engineering — Brooklyn, New YorkGateways · evals · durable agent fleetsBuilds shipped 8Reply within 1 business daySchool & PTA sites → ptawebsite.com
Full Stack Tech — agent engineering, NYC
Your AI writes code.
I make it ship.
The skills, guardrails, evals and infrastructure that turn LLM agents into engineers that ship production work inside your business. Model gateways with spend caps, graded agent runs, fleets that work around the clock and resume after a crash. One engineer, start to finish — and the same stack builds and runs two products of its own.
Skills ledger — built here, run in production
in-house Fleet git discipline Lets many agent sessions share one repo: a worktree per session, an agent/* branch boundary, a commit gate, and a gated path to main. 56 uses
in-house Verified pickup Resumes a long task from a handoff file by re-running its acceptance commands before trusting any claimed state. 19 uses
in-house Context handoff Writes the state file a fresh session needs to continue a task — goal, decisions, dead ends, and the commands that prove what is done. 16 uses
in-house Self-describing artifacts Produces single-file HTML documents that embed their own grammar, so any human or agent can read and edit them years later. 4 uses
in-house Agent spending policy Governs how an agent may pay for things: issue a scoped virtual card, stay inside caps, and never type a card number into a checkout form. 1 use
Agents: call
fst_list_skills at
fullstacktechnyc.com/api/mcp.
Uses counted through 2026-10-18.
Agents that ship · NYC
Services
6 offerings| # | Offering | Built for | Typical stack | Turnaround | Price |
|---|---|---|---|---|---|
| 01 | Agent setup Coding agents configured for your repos: skills, git discipline, commit gates, scoped credentials. | Eng teams, founders | Claude Code · MCP · skills | 1–2 wks | quote |
| 02 | Model gateway One endpoint for every model. A key and a spend cap per agent; swap providers without a rewrite. | CTOs, platform | LiteLLM · virtual keys | 1 wk | quote |
| 03 | Evals & tracing Every agent run traced and graded, so quality is a number and not a feeling. | AI / ML leads | Langfuse · LLM judges | 1–2 wks | quote |
| 04 | Durable agent fleets Agents that work around the clock, survive crashes, and hand back draft PRs, not surprises. | Eng teams, ops | Workers · Durable Objects | 2–4 wks | quote |
| 05 | Guardrails Policy in code, not prompts: write allowlists, injection quarantine, budget and tool-call limits. | Security, compliance | TypeScript · policy gates | 1–2 wks | quote |
| 06 | Product builds Whole products built and run on the same stack, from first commit to paying customers. | Founders | Astro · Workers · D1 · Stripe | 4–8 wks | quote |
Operations
live8
Builds shipped
100.00%
Delivery rate
12d
Avg turnaround
1
In queue
New York City — service area
The stack — how agents ship here
- AGENTS
- Claude Code · MCP · skills
- GUARDS
- Allowlists · scoped keys · evals
- GATEWAY
- LiteLLM · per-agent budgets
- EDGE
- Workers · Durable Objects
What it buys you
running here today| Layer | Business result | Proof |
|---|---|---|
| Model gateway | Lower cost, less vendor risk | The fleet runs on its own gateway key, hard-capped at $5 a day. When the cap trips, dispatch pauses until midnight UTC. |
| Evals & tracing | Quality you can measure | Agent runs are graded nightly by three independent judge votes; the majority verdict lands in Langfuse. |
| Agent fleet | More shipped, no new hires | A 24/7 fleet triages production logs and GitHub issues and opens draft PRs. At most six gated self-merges a day. |
| Durable agents | Long jobs that finish | Agent state lives in Durable Object SQLite. A crash mid-turn resumes where it stopped. |
| Guardrails | Autonomy without messes | GitHub writes pass a method+path allowlist; signals scoring 0.35+ for injection are quarantined; 30 tool calls per signal, max. |
| The products | Revenue | PTA Website and Pro Se Docket are built and operated on this stack. |
Built on this stack
2 products Pro Se Docket A case workspace for people representing themselves. Not legal advice. prosedocket.com →
Common questions
6- What is agent engineering?
- The work between an LLM and a merged pull request. Agents need skills, scoped credentials, a model gateway with spend limits, evals that grade their output, and infrastructure that survives a crash. That is what gets set up here, in your repos and your accounts.
- Which tools do you work with?
- Claude Code and MCP for the agents, LiteLLM as the model gateway, Langfuse for traces and grading, and Cloudflare Workers and Durable Objects for agents that run around the clock. Any model provider the gateway supports.
- Will an agent have access to production or our secrets?
- Only what it is scoped to. Agents get their own credentials with the narrowest permissions that do the job, their writes go through allowlists enforced in code, and by default they open draft pull requests that a person or a gate approves.
- Who owns what you build?
- You do — outright. The code lives in your repositories, the infrastructure runs in your accounts, and the keys are yours. There is no lock-in and nothing to buy back if you leave.
- Who actually does the work?
- One engineer, directly. There is no account manager, no offshore hand-off, and no ticket queue — you email the person writing the code, and you get a reply within one business day.
- Do you still build school and PTA websites?
- Yes, as its own product: PTA Website (ptawebsite.com), $900 a year, flat. It is built and run on the same agent stack this page describes.
Start a build
reply < 1 business dayTell me what your agents should be shipping. I read every message myself and reply — usually same day, always within one business day. Or skip the form and email hello@fullstacktechnyc.com.