Vinícius Raposo

Applied AI Reliability Engineer

AI systems you can leave on without babysitting.

Profile pin · Python

aria-agent

Agent harness whose approvals bind the exact action, with execution tracing and a Claude Code setup fixture (permissions, blocking hooks, independent verifier).

360 Python + 64 frontend tests; checksum-verified offline evidence. Native Claude Code trials have not been run, so no result is claimed for them; live PostgreSQL/Redis is not verified.

Language: Python Updated: 2026-08-15 Visibility: public
View on GitHub Back to catalog

Without a bounded harness, agent runs expand unchecked—tools fire without approval gates, skills leak context into every turn, and failures cannot be traced or replayed after the fact.

The harness bounds the model; skills disclose only when activated.

Offline demo:

cd frontend && npm run demo:ui

Walkthrough

Prompt injection blocked by deterministic safety before any tool runs.
Prompt injection blocked by deterministic safety before any tool runs. Sample data.
Safe calculator tool call with ordered execution trace.
Safe calculator tool call with ordered execution trace. Sample data.
Gated email draft waits in the approval queue.
Gated email draft waits in the approval queue. Sample data.
Skills disclose only when activated on a run.
Skills disclose only when activated on a run. Sample data.

Why it matters

ARIA turns approval, safety, trace, cost, replay, and reproducibility boundaries into inspectable local code instead of relying on a hosted control plane.

Verified proof

360 Python + 64 frontend tests; checksum-verified offline evidence. Native Claude Code trials have not been run, so no result is claimed for them; live PostgreSQL/Redis is not verified.