No-Code Agentist

Skills are the unit of growth. My new article is on the decision engine that made my Muse agent stop guessing and start knowing. https://t.co/Mu0K2vU5jE

Your dev team ships PRs in 10 minutes, but back-office takes 4 days just to clear an invoice. That latency gap quietly kills momentum. We've been fixing this at Deel with @get_akai A few practical things we learned when setting up these workflows: • Don't write custom glue code for every edge case. Just pass raw docs or emails straight to an agent parser. • Avoid giant mega-bots. Single-purpose agents with tight context perform way better and hallucinate less. • Always keep a human control layer. If an agent is about to spend money, update a system of record, or email a customer, require a quick human sign-off first. • Once approved, let it run in the background 24/7. We spent the last decade building APIs to connect software. Now we're using agents to handle the tedious logic between them. Deel even backs the implementation with a guarantee, so there's really no downside to testing it. Check it out: https://t.co/g4mqL6n5BO

When an approval request times out, I let it die. I do not let it turn into a send. Most automations do the opposite. No answer by the deadline reads as a green light inside most of these workflows, and the customer email goes out the door with nobody's name on the decision that sent it. Most workflow tools ship with continue on timeout as the easy default, because a stalled process looks broken and a moving one looks fine. That default runs backwards. A stalled approval is safe. A workflow that quietly finishes a job nobody signed off on is the real failure. My default on every timeout is dead-letter, not send. I escalate once if the deal is time-sensitive. I never let a timeout continue on into Gmail as though silence were an answer. Here are the actual clocks I run, all in Eastern time. A one-to-one email or Slack message gets about four business hours before it dead-letters. If the lead is already mid-conversation when that clock runs out, I escalate once to a named backup person for two more hours, then let it dead-letter anyway. A sequence, a quote, or an SMS blast gets twenty-four hours and exactly one named approver. It never auto escalates to a whole group, because a group chat is not an owner. The two clocks differ on purpose. A one-to-one thread goes cold fast, so four hours is already a long wait on the other side. A blast reaches many people at once, so it earns a slower clock and exactly one owner instead of a fast, trigger-happy one. I only escalate once on either clock, never on a loop. A second and third nudge just trains people to skim the message and clear it, the same habit Anthropic found inside Claude Code. The wait clock does not start until eight in the morning local time either. A card that quietly expires at three in the morning never trains anyone to check anything the next day. There are only four outcomes for a request like this: approved, rejected, escalated, or dead-letter. Exactly one of those four sends anything to the customer. A request that waits forever is not a safety feature. It is a stuck invoice nobody is chasing down. And if the timeout output feeds into the same Gmail step as the approval output, the system just built itself an automatic send with extra steps bolted on the front. Anthropic published a study on May 25, 2026, on how people actually use per-step approval prompts inside Claude Code. Roughly 93% of those prompts got approved, and the more prompts a person saw, the less carefully they read each one. Their fix was fewer, better-scoped asks, not more toast notifications. I apply that same finding to everything I automate outbound. The wire is whatever actually leaves the building: the email that reaches the customer, the refund that hits their card, the order that changes inside the store. That is what I gate. I do not gate the small lookups sitting in front of it, and a refund under the threshold I already trust does not need a person either, while one above it still needs a name attached before it fires. Fetching the last three emails or pulling up a record in the CRM should never stop and wait on a person. Ask a human on those small hops too and you rebuild the same 93% rubber stamp Anthropic already measured, just with more clicks along the way. I still review the dead-letter pile once a week. Most of it stays dead. A few need a person to finish what the agent could not. Silence still is not a yes. I let the card expire, keep the draft sitting in the queue, and send nothing until a named person approves it first.