Multi-Agent Workflows - what happened across three cohort sessions and one cross-cohort lunch and learn the week of April 7.
The most important thing this module was trying to say is that you are already doing multi-agent work. When you draft something in one model, take the critique to another, and iterate from there, you have a relay in motion. The sessions this week were about making that visible, naming its parts, and pointing toward where the architecture can go as your confidence grows.
The patterns covered move from simple to complex: a manual relay between a creator model and a critic model, an orchestrated team where one model coordinates several specialized agents, a debate structure where opposing models surface what your own thinking would miss, a shared data repository accessed by agents with distinct roles, and a fully autonomous pipeline where the human sits outside the loop. That last one is conceptual for most of us. The first three are not. Participants across all three cohorts were already running versions of them.
The practical takeaway: start with the relay. Get the creator-critic pattern into regular workflow. The orchestration complexity follows naturally once the basic habit is established.
The week's most instructive moment was not in any of the three cohort sessions. It was Friday morning, when Vinnie walked all three cohorts through the operating system he has built on his iPad over the past several months. He introduced it as the Etsy version of AI compared to Thor's Amazon - meaning handbuilt, personal, and genuinely operational rather than commercial-grade.
The system connects ChatGPT and Claude to Gmail, Google Calendar, HubSpot CRM, and Google Drive. It runs a morning check-in via voice dictation. Before each meeting it pulls email history, calendar data, and CRM context and produces a dashboard. Tasks complete with a toggle. The whole thing runs through voice on a second device so deep work stays uninterrupted on the main machine.
The origin matters: Vinnie built it first for a colleague with a physical limitation that made typing slow and painful. That colleague was being marked down on performance reviews for incomplete activity logs, not for incomplete activity. A voice-based system solved that. The personal OS grew from there.
His finding after running both platforms side by side: Claude has better connectivity and produces better dashboard outputs. ChatGPT is better for open conversation and some creative tasks. He uses both intentionally. That is calibrated, earned platform judgment built through real use.
Every architecture pattern covered in Module 6 appeared in the live demonstration. The morning check-in is a relay. The meeting dashboard is a shared repository accessed by a coordination layer. The voice-to-task loop is a manual automation. He did not build it by studying the architecture slides. He built it by solving a real problem incrementally, and the architecture emerged from the work.
The closing principle from the session: build incrementally, ask Claude to explain what it is doing as you build, and use the tool to teach you about the system you just built with its help.
The relay that went to a live demo. One participant ran his company's value proposition through three models in sequence: Claude for substantive critique, ChatGPT for comparison, Gemini as a third pass. Claude identified the existing tagline as architecture speak and gave concrete alternatives. The Claude-generated messaging went into a live prospect demo the next day, and his chief product officer called it out as a highlight. That is not a classroom exercise.
The single-tool constraint, handled correctly. A participant starting a new role this week learned her employer restricts AI use to Claude only. The group's answer was direct: use your own equipment and your own time to stay curious about the full model landscape. When you are inside the single-tool environment, you know what you are getting and what you are not. That informed perspective is a form of fluency that colleagues working only inside the mandated tool will not have.
The PDF extraction test. A participant ran a controlled comparison between Claude and ChatGPT using complex PDF reports with diagonal-label charts. Claude reasoned through the ambiguity, produced a clean spreadsheet, and flagged where chart formatting had likely introduced directional bias in its readings. ChatGPT misread monthly data as annual, insisted the monthly data did not exist, and then offered four complex workarounds for a problem it had created. The conclusion was unambiguous, and it produced a technique worth keeping.
The first autonomous agent, scoped correctly. A participant who manages a quality engineering organization described how her team chose their first autonomous agent task: test failure triage. The trigger is easy to identify, failure is easy to detect, and getting it wrong carries low consequence compared to autonomous code commits. Agents run. If a test fails, the agent kicks in. Before any code gets committed, a human does a PR review. That sequencing is the practical judgment this module was trying to develop.
The go-to-market prompt, built from the inside out. One participant wanted a better prompt for a go-to-market strategy but did not know where to start. He gave Claude a generic input and asked it to help him build a better prompt through conversation. Claude asked clarifying questions about audience, output format, and use case before producing anything. ChatGPT, run in parallel, rendered the prompt visually and handed it over early. When he ran the Claude-built prompt in a fresh session, his word for the result was "scary accurate." He arrived at something this module was saying from a different direction: the more time you invest in the setup conversation, the more efficient the work that follows.
The question that ended the week well. At the close of Friday's session, one participant described using Claude at 1 AM to find words for the parents of a 16-year-old who had died. He was stuck. Claude gave him a start. He made it his own. The model did not replace the human act. It unblocked it. That distinction has been the through-line of this program since Module 1, and it arrived without being taught.
Creator-critic-rebuild. Prompt a model for its best work. Instruct it to critique that work as a hostile competitor finding every weakness. Take both and ask it to rebuild from scratch incorporating the objections. The rebuild step is what most people skip, and it is where the real improvement lives. Documented in the Field Guide, credited to Josh.
Autonomous agent scoping heuristic. When deciding what to hand to an autonomous agent first, pick the task where the trigger is easy to identify, failure is easy to detect, and getting it wrong carries low consequence. Start there. Let performance build the case for expanding scope. Documented in the Field Guide, credited to Michele.
Architecture visualization. Ask Claude to display the architecture of your workflow or configuration as an HTML output. This produces a visual diagram of what the system pulls from, how it connects, and what it produces. Complexity that felt manageable in prose becomes visible as a diagram. Documented in the Field Guide, credited to Vinnie.
PDF bias spot-check. When Claude extracts data from PDFs with diagonal or rotated labels, ask it to flag where label angles may have introduced directional bias. Then spot-check those values. This converts a silent error into a flagged and manageable one. Documented in the Field Guide, credited to Eric.
Updated through Module 6. Four new entries from this week's sessions. If anything is misattributed, send a note.
Several participants noticed Claude and other platforms felt slower over the past week or two. This is real. Rapid adoption is outpacing available compute capacity, and Anthropic has been adjusting model availability and context limits in response. The practical implication: intentional model selection matters more now than it did three months ago. Using Opus for administrative tasks when Haiku would do the same work is a habit worth correcting. The model selection page from Module 3 is still the right reference.
The email triage system demonstrated across this week's sessions is something actively being built, not something already finished. The config file exists. Roughly thirty build steps remain. The principle behind it is more important right now than the finished artifact: intentional human friction before anything goes out. The model drafts. You review. You push send. That sequence does not change regardless of how good the drafts get.
You are at the end of Phase 2. The Phase 1 feedback reshaped Modules 4 through 6. Your Phase 2 feedback will shape the build and capstone phase and inform how the program is designed for the next cohort. It takes about five minutes.
Your feedback shapes what comes next. Takes about five minutes.
Before next week's session, use your AI of choice to sharpen your cohort's project statement to one specific, solvable problem. Ask it to poke holes in the scope. What are you solving? What are you explicitly not solving? Come to the session ready to say that in two sentences.
Start the email thread with your cohort. Use the contact information from your meeting invites, reply-all, and begin the coordination conversation. By the time you arrive at next week's session, your group should have a working direction, a rough role assignment, and a sense of who is doing what during the build sessions.
Next week is about definition, design, and delegation. Come ready for all three.