Aroha Labs

Blog

Engineering insights, protocol deep-dives, and stories from building the trust layer for AI agent networks.

Engineering6 min read

Your Agent's Permission Logic Lives in 2,267 Files

2,267

files, one rule

3

narrow waists

1

audit log

New research traced one safety rule — 'confirm before file deletion' — through a production agent harness and found it scattered across 2,267 files. There are two honest answers to that: better maps of the scatter, or moving the rule somewhere the scatter can't reach.

Read the post
Security7 min read

Your AI Agent's Tools Need a Firewall

--block

never happens

--gate

asks you first

--limit

hard budget

Your assistant's GitHub server exposes delete_repository. The only thing between the model and that call is the model deciding not to. mcp-guard wraps any MCP server with block rules, human approval gates, and session budgets — one config line, zero code, fails closed.

Read the post
Tutorial10 min read

Five Packages That Changed How I Build AI Agents

A working tour of the Aroha toolkit in Python and TypeScript — a five-line agent, a firewall for MCP tools, spending limits that math enforces, delegation webs with receipt trees, and live agents you can call today. Written by the person who built it, bugs included.

Read the tour
Security8 min read

The $47,000 Wake-Up Call: Why AI Agents Need a Permission Layer

$47,000

billed

0

auths

6

sub-agents

A startup gave their AI orchestrator access to a company API key and told it to research competitor pricing. By Monday morning it had generated $47,000 in charges — and nobody had authorised any of it. Here's why existing tools can't prevent this, and what actually needs to happen.

Read the story
Protocol7 min read

Spending Mandates vs OAuth Scopes: Why AI Agents Need Different Authorization

OAuth was designed for humans granting access to apps. When one AI agent delegates to another — and that agent can delegate further — OAuth scopes break down fast. Spending mandates solve three problems OAuth never had to think about: attenuation, chain depth, and spend limits.

Read the post
Security6 min read

Five Reasons Your AI Agent Should Never Hold an API Key

API keys are all-or-nothing, indefinitely-lived, and completely silent about what they were used for. Giving one to an AI agent is like handing a contractor a master key to your building — no entry log, no expiry, no way to limit which rooms they can open.

Read the post
Engineering9 min read

How We Built a Reputation Engine That Gets Smarter with Every Call

The Aroha reputation system uses a Bayesian Beta distribution — not star ratings — to score agents. This means an agent with 3 perfect calls beats one with 1,000 mediocre ones at a statistically correct threshold, and the uncertainty shrinks with evidence rather than with time.

Read the post
Tutorial10 min read

From Zero to Production: Deploying Your First Aroha Agent on Fly.io

A step-by-step walkthrough: scaffold an agent with npx create-aroha-agent, register it on the Hub, issue a spending mandate so it can call a downstream API, and deploy it to Fly.io — all in under 20 minutes. Includes the Dockerfile and the exact fly.toml config.

Read the tutorial

More posts on multi-agent orchestration, DID key management, and CrewAI integration coming soon.