Internal AI tooling
Agent systems we built and run
Published September 17, 2026
The problem
Every engineering firm currently claims AI capability. Almost none can show what that means in practice, because for most it means calling an LLM API inside a client feature and nothing more.
There is also a real internal problem. A firm doing specialised technical work has to maintain visibility across several platforms, each with its own conventions and limits. Done manually, it is a recurring tax on the most expensive time in the business.
What we built
Agentic publisher
An internal system on MCP plus direct LLM calls: it ingests a source, drafts platform-specific variants rather than reposting one piece of text everywhere, schedules against a calendar, and publishes via API to X, LinkedIn and Telegram. Reddit is planned.
Safety hooks
On our platform work, guardrails that block high-value transactions without explicit confirmation and prevent accidental exposure of key material, a materially different exercise from making a chatbot answer questions.
Background agents
Parallel multi-chain deployments and long-running cryptographic ceremony monitoring.
Automated review
Agent tooling running against our own repositories as part of normal development.
Outcome
- A repeated internal task moved from manual work to a supervised agent pipeline
- Agent tooling embedded in our own delivery, not only in client deliverables
- Safety guardrails built for genuinely dangerous operations, high-value transactions and key material
- We use what we build. When we specify an agent system for a client, we are describing something we operate
