Platform | GRAIsolSkip to main content

The platform

Agent AFK is how I ship the work.

A self-hosted agent runtime built from scratch on Anthropic's raw Messages API — CLI, daemon, and Telegram bot sharing one config and a library of built-in orchestration skills. Not a wrapper. Not a framework. The actual harness that runs every engagement.

What it does

Agents that plan, build, verify — without supervision.

The agent loop is mine. Built directly on Anthropic's Messages API, not wrapped around someone else's harness — session lifecycle, tool dispatch, hook firing, sub-agent spawning, cancellation semantics, prompt caching, and persistent state, all engineered so an agent can run for hours and still know what it's doing.

Slash commands compose real pipelines. /mint takes an idea and ships a verified implementation end-to-end. /diagnose forks parallel hypothesis sub-agents to root-cause a failing test. /forge proposes new skills and gates them behind L1 capability evals before they ship. Every workflow runs on the same orchestration core.

I built it because I needed it. Client engagements, internal automation, outbound systems — all of it runs through the same harness. Same decision loop, same audit trail, same recovery paths. The platform that runs my own work runs every engagement too.

Why it works

Built different on purpose.

01

Multi-Surface, One Runtime

CLI, background daemon, Telegram bot — three surfaces, one session manager, one config root. Start work at the terminal, get the result on your phone. The runtime doesn't care which door you came in through.

02

Built-In Orchestration Skills

Slash-commands wired to real pipelines: /mint runs the full spec → research → build → verify loop. /diagnose forks parallel hypothesis sub-agents. /forge generates new skills against a capability gate. Composable, not a wrapper.

03

Telemetry-Driven, Self-Improving

Every sub-agent dispatch appends to a JSONL log. Patterns get distilled into pattern cards, scored against falsifiers, and promoted into new skills behind a qualify gate. The runtime gets sharper every week without anyone deciding to make it sharper.

04

Engineered for Long-Running Work

AbortGraph for clean transitive cancellation across sub-agent trees. Dual-provider across Anthropic and OpenAI Codex. Per-symbol audit of every SDK surface used, gated in CI so upstream drift fails the build instead of corrupting a four-hour run. Boring infrastructure for an unboring runtime.

The stack

Boring tech, sharp choices.

Runtime

Anthropic Messages API · OpenAI Codex SDK · TypeScript strict · Node 20+

State

SQLite (better-sqlite3) · JSONL telemetry · cross-session memory

Surfaces

Interactive CLI · headless daemon · Telegram bot (Telegraf)

Distribution

Self-hosted · BYO API key · git-based plugin system

Have something gnarly to ship?

I take on a small number of founder projects each month. If you're past “we need AI somewhere” and ready to put something in production, let's talk.