Agents we’ve built
Prototypes, internal tools, and open source. Each is labeled for what it is, and none is presented as a client deployment. Test counts are from our own runs of each project’s suite.
Prototype
KOTA
Phone ordering agent, English and Mandarin
Voice, typed tool calls against a database, confirmations that are checked, and a hand-off to staff.
What it does
Answers a restaurant’s phone, takes the order against the restaurant’s own menu database, saves it as a structured record, and texts the caller a confirmation. Pressing 0 transfers the call to a person.
How it’s built
Twilio media streams; Deepgram for English; ByteDance speech with OpenAI tool calling for Mandarin; Python and FastAPI; PostgreSQL.
Evidence
111 automated tests pass. Built against real restaurant menus; not used in production.
Internal tool
Research agent with source checks
Research and monitoring
An agent whose citations can be trusted: it only counts sources it actually read.
What it does
Runs web research, records every source it consulted, and grades each one. A citation counts as verified only if the agent consulted that exact URL during the run, and a finding needs at least three independent, credible sources. A person must approve before anything is spent, sent, or published.
How it’s built
Node.js, OpenAI web search with structured outputs, SQLite for runs, sources, evidence, and approvals.
Evidence
208 automated tests pass.
Internal tool
Approval-gated outreach pipeline
Lead research and follow-up drafts
AI drafting with rule checks and a person in the loop before anything leaves.
What it does
Takes a lead list, checks the opt-out list, researches each business from its public site, and drafts a first message. A rule check rejects claims the product can’t back. A person approves every message, and the send step enforces a daily limit and refuses duplicates.
How it’s built
Python, SQLite, and a choice of Anthropic, OpenAI, or Gemini models. Each draft is stored with a fingerprint of the capability list it was checked against.
Evidence
157 automated tests pass.
Open source
ARCHON
Agent runtime and orchestration
The plumbing behind agents: tool registries, worker delegation, MCP, memory, and approvals.
What it does
A personal agent runtime that runs from a terminal or Telegram. It registers tools, delegates coding tasks to Claude Code, Codex, or OpenCode workers, connects MCP servers, keeps memory across sessions, and blocks risky shell commands unless a person confirms them.
How it’s built
Python, with Anthropic, OpenAI, and Google model support.
Evidence
Open source and in active development. Experimental.
Need an agent like one of these?
Tell us what happens today and what a good result looks like.