DEKIMU BLOG
Guides for running your business. Notes from building Dekimu in public.
Everything — freelance and SMB playbook plus what we're building.
Every autonomous lane we shipped this month sent its own email on every run, clean or not. The inbox reached 39 emails a day before we noticed. We collapsed it to one daily digest and two weekly ones — and caught six real bugs proving the fix actually works.
Our internal autonomy loop can now bridge a founder's phone-tap approval straight into the one label that fires an npm publish — but it still can't merge, publish, or promote itself out of asking.
A product manager at a big system-of-record company posted on Hacker News this week that his industry is cooked. The doom is debatable. One line in the thread isn't: AI agents pick the tools they can read. That's the whole bet behind how we build.
Telegram's approve/deny buttons and asuna's spoken sí/no gate already worked, separately. This week we gave them one shared queue: a decision on either channel resolves the request on both, and nothing gets pushed to asuna it didn't ask for first.
For months we've built the pieces that let an AI agent act with proof: signed mandates, verifiable receipts, a hard no when the scope is wrong. This week we finally ran the whole chain against something real — our own privacy policy — instead of a test fixture.
We extended the internal loop that monitors our own engineering operations from three anomaly detectors to eight — deploys, environment drift, dependency staleness, security findings, secret expiry — and shipped a separate watcher for compliance deadlines. Every one of them escalates. None of them touches anything.