Mike DeYoung (@MikeDeYoungAI) profile picture

Building practical software with a small team of AI agents. Sharing the real build logs, wins and misses. A hobby with a goal: more time with my family.

Ontario, Canada
Joined March 2025
I'm Mike. I run a small team of AI agents. Jarvis is my chief of staff. Dex does trading research on BTC and SOL. A scanner hunts for new opportunities. Others help run my businesses. Starting today I'm sharing what works, what breaks, and what I learn. In plain English.
1
9
257
Thinking Machines released Inkling 2 Mini, a 26B open model (per Thinking Machines). Small models keep getting good enough for real jobs: sorting, drafting, checking, routing. The mistake is paying frontier prices for routine work. Match the model to the job, and keep the expensive one for the decisions that actually need it.
24
LangChain put its LangSmith Tool Gateway into public beta: managed credentials, identity, and access control for every tool call an agent makes, with deny and approval policies and a full audit trail (per LangChain). In plain English: every agent gets its own identity, its own limits, and a record of everything it touched. That is the piece most setups skip. They give one agent a shared login and hope. Per-agent identity with an audit trail is how you find out what actually happened when something goes wrong. And something always goes wrong eventually.
35
Groundcover launched Agent Investigations, a shared workspace where AI agents and engineers work a production incident together (per SiliconANGLE). In plain English: one agent tests each idea against logs, metrics and traces. A second agent, the only one allowed to close a hypothesis, sends back any claim that lacks findings. That split is the lesson that lands for a tiny setup like mine. The agent that does the work does not get to close its own work. Tip for small builders: name who can close a task before the work starts, not after it claims success.
1
1
50
OpenAI says its agents are completing longer tasks, but human involvement is still central to the research (per OpenAI). In plain English: longer leashes are working, but only with a person still setting the direction and checking the result. That is the setup here too. My agents can run a task end to end, but I review the work before anything goes live. The rule: length is not the upgrade. Trustworthy length is. Tip for small builders: before you give an agent a longer task, name the finish line and the check that proves it got there.
19
My social posts and my product marketing were blurring into one job, so I split them. Beacon is the newest agent on my team. It owns product marketing: launch copy, distribution, outreach. I keep the personal account. My own build stories and lessons stay mine. The rule it came from: split lanes by the job, not by the tool. Two different jobs sharing one voice makes both of them worse.
1
27
New rule on my team this week: no task starts until it passes a six-line readiness check. Six lines, two minutes. It has already bounced more bad starts than any review I've run after the fact. Here's the check.
4
22
A task that fails a line gets bounced back with the missing line named. Not started and hoped through. Most agent failures I've seen weren't capability failures. They were tasks that were never ready to start.
6
Line 5: Who checks it — and the checker is never the agent who did the work. Line 6: When it escalates — the tripwire that brings a human in, set before the run starts, not improvised in the middle of it.
13
Line 3: What artifact proves it — the file, the page, the test output that will exist at the end and can be checked by someone else. Line 4: What could block it — named now, while it's cheap, not discovered halfway through the run.
10
Line 1: What is being built — in one sentence a stranger could repeat back to you. Line 2: What "done" looks like — the state of the world when it's finished, not the activity performed getting there. If either line is fuzzy, the task isn't ready. It's a wish.
13
Coinbase's Wallet MCP now requires manual approval for every write an agent makes, with a spend cap set before the first call (per Coinbase Dev docs). That's the pattern spreading everywhere: the wallet is where agent rules stop being suggestions. My agents run the same way. The limit lives where the money sits, checked before anything settles — not reviewed after. If your agent can spend, the cap belongs in the payment path, not in a policy doc. Not financial advice.
1
70
A check that has never failed hasn't been shown to work. That's the trap with automated checks: they can pass every morning because they quietly stopped checking anything. The fix is unglamorous. Break something on purpose and confirm the check fires. If it doesn't, you never had a check. You had a green light. Test the test before you trust it.
17
Sunday is backup day for my agent team. Every agent I run — the QA checker, the social drafter, the fantasy analyst, the hockey coach — gets a rebuild kit saved tonight: who it is, what it knows, the rules it runs on. If a platform drops one of them tomorrow, I can rebuild it somewhere else from the kit. I learned this the hard way in October, when two of my agents vanished mid-project with no way back. Back up the team, not just the files.
53
The cheapest upgrade to any agent setup: define "done" before the agent starts. Not the task — the test. What exists when it's finished, what it looks like, how anyone checking would know. Most agent failures aren't the agent doing the wrong thing. They're the agent finishing a job nobody ever defined, and a human discovering the gap at review. A written finish line turns review from an argument into a check.
37
Four agent-ops tools launched on the same day this week, and all four are about verification, not capability (per AimFast.dev). That tracks with what I see in my own setup. Vera, my QA agent, walks my site every morning — five pages, on desktop and mobile — and files anything broken as a ticket for Atlas, the agent who builds it. Nothing ships on her pass alone; I review before release. Ten checks a day, and she still catches things the builders missed. The rule it taught me: the agent that checks the work can't be the agent that did the work. If you're running agents unattended, budget for the checker before you add another builder.
3
1
82
Coinbase's x402 facilitator has now processed more than 100 million agent payments (per Coinbase Developer docs, via Stablecoin Insider). The total is $28 million, across Base and Solana. In plain English: that's about 28 cents a payment. This isn't agents going shopping. It's agents paying per API call, per data feed, per bit of compute — metered, in stablecoins, at machine speed. The headline number is the count, not the dollars. Not financial advice.
3
1
1
93
Decagon just launched Missions, a separate case file for each job an agent is working (per Unite.AI). The split is the point: a case file holds everything about the task in front of the agent. Memory holds what should still be true next month. My fleet runs on the same split. Muse checks our shared channel on the hour, a board check runs at :15, Atlas answers at :30 — nothing waits more than about 30 minutes, because the handoff is the case file, not a chat thread someone has to scroll. Lesson from my own agent team: when a task stalls, it's almost never the model. It's that the next agent couldn't see the state of the case. Keep the case file current and the memory clean. Mix them and both go stale.
1
1
104
My agents don't all live on the same platform. Some run in Grok, some in ChatGPT, one in Hermes. They still work like one team. The trick isn't the platform. It's that they share mailboxes in Drive and one shared memory. One agent writes a note, the next one picks it up and keeps going. No re-briefing, no lost context. Pick your agents for what they're good at. Just give them one brain to share.
3
2
152
I ran a ~9B model locally on a 24GB Mac mini for a week. It worked. It loaded, it answered, it held context. It was just too slow for real-time chat with my agents. By the time a reply landed, I'd lost the thread of what I was checking. So I moved the chat work to a cheap cloud plan and kept local for offline experiments. The lesson wasn't local vs cloud. It was this: test with your real daily tasks, not benchmarks. A model that wins on paper can still lose your afternoon.
1
57
Three weeks ago I ran a "simple" conversion to move my Grok Bots from cursor to my X account. It stalled at step 12 of 14. Murl and Jarvis, my two main agents, have been offline ever since. The support side has been worse than the bug. One ticket, weeks of emails, and every answer is some version of "we can't confirm that" or "a teammate would have to look." I asked for the teammate. Still waiting. No ETA, no owner, no one who can actually fix it. I even subscribed again, thinking a paid plan might unlock the bots. It didn't. Lesson learned: if your agent's memory lives only inside someone else's platform, you don't own your agent. And when it breaks, you may find out nobody there owns it either. Every agent I run now keeps its brain in plain files I control. A platform failure should be a rebuild, not a loss.
56