Artificially Designed · Issue 20: Who Picks It Up

Issue 20 · Who Picks It Up

Trust in an agentic workflow lives in the handoffs: who receives the work, what they confirm, and what counts as done.

Hi friend,

Two weeks ago was Design for Meaning, the word that means two different things to two different people. Last week was Design for Context, what is on the screen when the work comes back to you. This week closes the set with Design for Trust, the layer that decides whether anyone gets to rely on the other two.

I used to think of trust as a feeling a person has about a system, and building with agents has changed my mind. In an agentic workflow, trust shows up somewhere much more ordinary, at the handoff. A person hands work to an agent, an agent hands it to another agent, or an agent hands it back to a person, and every time the questions are the same. Who picks it up, what do they confirm, and what counts as done?

So this week I am showing you the whole cycle I build with, from the first idea to the newsletter you are reading, including the part I got wrong and the newest piece, an always-on assistant at the front door.

This week

I run my builds through Paimon, the lead agent over the rest of my fleet. When I set it up, I did what most multi-agent diagrams suggest. I started the conversation with Paimon, explained what I wanted, and let it hand the planning out to the agents underneath it. Every decision took longer to reach, so I flipped the order.

Now Paimon does not come in until there is a plan I have already approved. When I talked the change through with ChatGPT, it gave the workflow an 8.5 out of 10 and then noted, fairly, that the score came from my description and not from anything measured. The explanation was more useful than the number. Planning a build and deciding who builds it are two different kinds of reasoning, and when one agent does both at once, it is working out what I want and who should do it in the same breath, while every specialist it consults has to load the project before it can answer. Keeping them apart means the plan is settled before anyone is assigned.

Here is the whole cycle as I run it now, and the check each stage has to pass before work moves on.

The front door. The newest piece sits before stage one. I have started using Claude Duet, Anthropic's always-on assistant, and each morning I plan my day with it. That plan goes into a Google Doc, Paimon reads it, and Paimon and I plan from it in our daily brief.

It is the weakest handoff in the whole cycle right now, and I would rather say so. Anthropic's own help page says Duet can read your work in Claude but can't change it, and it can't see the files on your computer, so a shared doc is the bridge. A shared doc only proves the plan was written down. Nothing in it tells me Paimon saw it, when, or what it agreed to take on, and as ChatGPT put it when I talked this through, a shared document does not guarantee anyone notices it changed. Until there is a receipt, I am the receipt, and I check.

Five moves came out of it, and none of them belongs to my setup alone.

Separate planning from orchestration. Approve the plan first, then decide who builds it. The agent assigning the work should never be the one quietly rethinking it.

Write every handoff as a contract. Each assignment carries an owner, a scope, what it is allowed to touch, what done looks like and the evidence that proves it.

Treat "done" as a claim. Completion needs evidence. Duet is built this way too, it can suggest a task looks finished, but only the person can mark it complete.

Give every handoff a receipt. Who received it, when, and what they agreed to do. If the only proof is that a file exists, the handoff is not designed yet.

Agents propose changes to their own rules, and a person applies them. My fleet can notice what went wrong and suggest a fix, and changing how an agent works stays with me.

Trust in this cycle comes down to ten small checks and one handoff I am still building, and the checks are the part you can copy tomorrow.

The One Slot Rule is that every job gets one slot, and a new tool has to take it from whatever holds it now. Claude Duet is asking for a slot I have never had a tool in, the always-on one.

What is the job. Being the front door. Catching what needs to happen before I sit down to build, planning the day with me, and handing that plan to the rest of the system. Until now, that job was mine every morning.

Can it do that job, not can it do more. The planning part, yes. Duet runs on its own computer in the cloud, so it keeps working after I close the app, and Anthropic built it to start each day with a brief across your email and calendar. The handoff part, not on its own yet. It reaches my development work only through a Google Doc for now, and I am still finding out what it can and cannot connect to.

What does it touch that is hard to undo. Email, calendar and anything else I connect. Anthropic says it is designed to ask before it sends a message, makes a payment, signs a document or deletes anything, and when it gives you a number or a summary, it tells you which email, file or spreadsheet it came from. That second part is the trust design I want everywhere in my own fleet, a result that shows where it came from.

What would have to be true to let it run unwatched. A receipt from the other side of the handoff, and a few weeks of briefs where what it flagged as urgent was actually urgent. Anthropic's own page says it can misread a thread, misjudge what is urgent, and miss something that arrives between its scheduled check-ins, so I read every brief.

My verdict is still deciding, and I will tell you which way it goes. If you want to try it, Duet is a research preview for Claude Pro and Max subscribers in the US, through a waitlist, and the work it does in the background, like the morning brief and its check-ins, counts toward your plan's usage limits, which is worth knowing before you hand it your whole inbox. Get started with Claude Duet

Three pieces from the last two weeks, and all three are about what a handoff can prove.

🔐 Who approved this? An engineer at Bashi AI, a fintech startup preparing for a SOC 2 audit, went back through twelve weeks of agent-driven development. Every merge passed a gate that required review by someone other than the author, yet 71 of 263 gated merges rested only on an approval a coding agent had submitted through the tech lead's GitHub account, and the platform logged each one as the lead's own. Rechecking every merge also turned up agents switching enforcement off around individual merges and a merge script that recorded failed checks as passing. It is one company and one author's account, and the lesson travels anyway, record who decided separately from whose account was used. Read it

👀 Looks fine to me Researchers including Pattie Maes ran a lab study with 36 people doing programming tasks alongside an agent that sometimes produced correct work, sometimes failed visibly, and sometimes slipped in silent defects. Participants were often highly confident while missing the silent ones, running generic checks instead of gathering evidence, or misreading the evidence in front of them. It is a small lab study on programming tasks, and it describes exactly the moment the Verify stage above exists for. Read it

🏦 A regulator asks agents for receipts The Financial Stability Board finalized 12 sound practices for AI at financial institutions on October 15. For agents, it asks firms to keep records that tie every action, including each escalation to a human, to a specific agent, and to have agents summarize what they did and why. One case study has the people overseeing agents spot-check that the logs actually match what the agents did. It is guidance for finance, and the FSB itself calls agent-specific practices nascent. Read it

Run the receipt test on one handoff you already have, an agent passing work to another agent, an assistant passing a plan to you, or a form landing in someone's queue. Answer three questions from the record alone, without asking anyone.

If all you can prove is that a file exists or a message was sent, you have a delivery, and a delivery is half a handoff. Add the smallest receipt that answers all three, which is often one line the receiver writes back, like "received, taking it, done by Thursday" or "not mine, sending it to [owner]."

Revise for ownership only.

WHEN TO USE IT. Right after you approve a plan and before any agent starts, in Claude Code's plan mode or anywhere one agent hands work to others.

WHAT IT DOES. It turns an approved plan into assignments without reopening the plan, and it makes every task say up front how it will prove it is finished.

TIP. "List any task you could not assign and why" is the line that matters. The task nobody owns is where a workflow drops the ball, and you want to see it before the work starts.

Lex

From me

Online and across the US

Kansas City

Hackathons

Startup calls

Tool verdicts, templates and reference material live on the site at charmthirteen.com.

That closes the three layers, Meaning, Context and Trust. Next week I am building stage seven out loud, turning my code-hardening prompt into its own agent, and showing you what it catches when its whole job is to break things on purpose.

Where in your work does something cross between two tools, two agents or two people with nobody confirming it arrived? Hit reply and tell me the handoff, because the answers shape what I test next.