The Agent Builder Loop from Daily Work to OSS

AGNTCon + MCPCon Japan
https://events.linuxfoundation.org/agntcon-mcpcon-japan/

Transcript

1. The Agent Builder Loop from Daily Work to OSS Minoru Onda @minorun365
2. Minoru Onda @minorun365 Technology Evangelist KDDI Agile Development Center Corporation AWS AI Hero / AWS Samurai Qiita 2025 Top Contributor 2
3. By the way, I was on TV last night The archive is on TVer, free to watch 3
4. Who here uses a coding agent?
5. Who uses it for something other than code?
6. My work is more than coding Expenses, attendance, travel Talks, slides, coordination Presales and delivery Coding agents work beside me across all of it. 6
7. A coding agent is a talented new hire Not an RPA tool. Three sharp consultants, full remote, assigned only to you. But it is their first day here. 7
8. Example 1: Expense reports 1. Hand over the manual and the URL of the system 2. Say what I want, in one breath "Settle last month's Claude bill. Find the receipt in Gmail, download the card statement, black out the other lines, and fill in the form through the browser. Stop before Submit." 3. It starts, and I move on to something else 8
9. Example 2: Project work 1. One folder per project: transcripts, decks, customer docs 2. Recurring jobs become skills: minutes, agenda, slides 3. Push to GitHub Enterprise. The team works in it too New member, no time to onboard? "Clone it and ask Claude" 9
10. Both examples have the same shape Asks when unsure Human Agent Human Goal and context Gather and shape Confirm I call this semi-automation. The middle is the agent's. The final call is mine. 10
11. Honestly, one task alone is not always faster Faster on a single job? Depends on the job. What did change: The work runs without me, so I do something else I hand out several different jobs at once 11
12. Three jobs run while I sit in a meeting I focus on the meeting In parallel Agents: direct and approve only at key points 12
13. How do you make its output actually good?
14. Never let it build the finished thing first I talk through the goal, the reader, and the message. It drafts the outline, and we argue until the shape is right. Quality Done 90% 0 Goal, reader, constraints Argue over the outline Let it write the body Fix what bothers me The last 10% is mine 14
15. Not happy with what it makes? Most people give it far too little
16. Talk to it out loud, and talk a lot What a prompt needs is volume, not polish. "Um", "no, wait", starting over: all fine. Talk the way you talk to a friend, and hand over the whole context. 16
17. Can you really trust it with your company's systems?
18. No MCP, no SaaS access? There is a trick It cannot touch the expense system? Then it prepares the input for me. "I will operate the expense system myself at the end. Put what goes into each form into a table I can copy and paste." Only a human touched the system. No credentials shared. 30 minutes became 5. 18
19. No need to fear everything you delegate Examples Reversible Drafts, reading, tables Irreversible Send, submit, delete, publish How I delegate Let it run A human confirms first When unsure: "If this fails, can I fix it on my own machine?" "Sign-in, send, confirm: always a human." Written in the rules, not remembered 19
20. What should an agent actually do?
21. My agents begin as work, not as product ideas Daily work Friction Build an agent Operate it Open source What people report becomes the next friction Every agent I run started as something that annoyed me during an ordinary week. 21
22. ① An agent that builds slide decks Free for anyone to use. 1,700+ registered users 22
23. ② An agent that watches public tenders Automates the hunt for public-sector system development projects Nine tender sites, twice a day New calls and deadlines to Slack 23
24. ③ An agent that registers community events Auto-registers new events to JAWS-UG's Twitter announcement form 24
25. So, how were these agents built?
26. Slide generation was my repeated pain Dozens of talks a year, every one from an empty deck. Same research, same layout calls, same export steps. 26
27. That is when I met Marp An OSS born in Japan. You write slides in Markdown, no PowerPoint needed. These slides are Marp too. 27
28. So I let Claude Code write the deck 28
29. Not everyone around me runs a coding agent So the same flow became the web app from ①. Sign in, describe the talk, download the deck. A semi-automation that worked for me, handed to people who cannot run an agent. 29
30. But building one is hard, right?
31. An agent is three lines of code now Strands Agents, for example from strands import Agent agent = Agent() agent("What does Strands mean?") Run it, and an agent is alive. 31
32. And deploying it serverless is just as easy AgentCore Runtime: containerize it, then just deploy. 32
33. The architecture AWS Cloud Serve the UI CloudFront Lambda Web Adapter User Build the slides AgentCore Runtime Strands Agents Marp CLI Tavily Sign-in Cognito Bedrock S3 33
34. Building is easy now. The bottleneck is use cases
35. I look for use cases in sales meetings With our own sales team, across many industries Every month, requests like this pile up: "We would like to build an AI system too, and make our operations more efficient." Vague. No use case yet. Customers cannot specify an agent they have never used. 35
36. Vibe Presales. Built while we are still talking A small agent, written during the meeting "Is this the thing you wanted?" Whoa!! We want this! 36
37. A demo changes the conversation Until they see one running, an agent is an imaginary creature that lives in PowerPoint. After the demo, they talk about data, approvals, edge cases, and who owns it. That is where the real requirements show up 37
38. Useful lessons should leave my laptop
39. Everything I build is open source I want people to look and think, "then I could build one myself." 39
40. Bugs I hit along the way go back upstream Not a workaround, but a pull request. 40
41. Inside the company, it spreads as InnerSource Write down the rules One win Add every stumble A small team One team, one template The way of working: rules, skills, and hooks. Clone it, and ask the agent how it works. 41
42. The code shows what. The why needs words Code, so anyone can inspect it Articles, for the decisions that hurt Books, for the whole story Someone clones it. Their Monday starts the loop again. 42
43. Why not start today, and hand your work to Claude Code?
44. Minoru Onda / Technology Evangelist KDDI Agile Development Center Corp. Amazon Pre-order Bestseller #1 Out September 24