Skeleton Crew

Sadi Kaya has a team working on his software. Every one of them can read code, judge it, and send it back. Only one is allowed to write it. None of them is a person.

Kaya is the solo founder of Indebtio, a Danish product for self-service debt collection. It's a live business with paying customers, built in the margins of his week. He runs the whole development cycle through a pipeline of agent personas: a refiner, a designer, an implementer, a reviewer, a QA agent, a security agent, and a legal jurist that checks Danish debt-collection statute before any code reaches him.

The agents open pull requests. They are never allowed to merge. "the routines never merge," Kaya writes. "they're not allowed to."

That one rule is the whole essay. Kaya keeps exactly two decisions for himself: what's worth building, and what's good enough to ship. Everything structured in between runs while he's away from the keyboard. "The leverage is in the middle," he writes, "and the judgment is at the ends."

An agent can tell you it's done. That isn't the same as the work being done.

Welcome back to Skeleton Crew. Let's get to work.

This week in small teams and big agents.

🔗 No agent grades its own homework

Maximilian Gutsche, solo founder of Unifire.ai, runs one workflow across four agents, and the rule holding it together is that no agent reviews its own code. He writes a single portable spec (scope, subtasks with acceptance criteria, risk-tagged assumptions, validation scenarios, target files) and hands it down the line: Claude Code plans, Cursor integrates, Copilot completes, Gemini reviews. "Don't let the implementor grade its own homework," he writes; the reviewer is always a different model than the author. When an agent drifts to working off its own last answer instead of the spec, he restarts the session with only the spec and current task. Write the spec as your contract before you touch an agent, and route review to a different model. [READ MORE]

🔗 Your monitoring misses the failures that look like successes

Arunkumar Molugu catalogs the agent failures that never throw an error: tools silently skipped, a state file claiming 13 items when the disk has 7, an agent answering a stock-price question without calling a single tool. He ran a known bug through his own detector, and it scored clean 100 out of 100, because the detector hunted for error keywords and the bug produced empty content with a status of "skipped." Stop gating on errors and non-200s. Check whether the tool actually got called, whether two consecutive outputs are identical, and whether final state matches what the trace claims. [READ MORE]

🔗 The 60% is easy. The 90% is the wall.

Austin Xu runs three agentic tracks at once: a spec-driven build pipeline, a support and triage line, and a management layer wired to JIRA, Gmail, and Docs. His self-reported experience: getting an agent to clear 60% of the operational grind is easy, 90% is dramatically harder, and 90 to 95 harder still, because what's left calls for judgment no runbook covers. "Files are the source of truth," he writes. Scope your next automation to a closed problem with a runbook you trust, and budget for the 60% ceiling before you commit to the climb. [READ MORE]

🔗 Degrade, don’t break

Kite AI co-founder Chi Zhang describes a Control Plane: a drop-in, OpenAI-compatible gateway that reroutes to a cheaper model at 90% of your spend budget and only hard-blocks at 100%, so a job finishes degraded instead of dying mid-task. This is an announcement, not a review. No pricing, nothing tested here. The pattern is worth stealing whether or not you ever touch the product: wire a fallback-to-cheaper-model trigger at around 80% of your monthly budget so jobs limp to the finish rather than failing silently. [READ MORE]

🔗 664 tools in context before the agent reads the task

Andrew Rudchuk, solo founder of Artifi, runs governed finance agents on real companies' books, and he did the token math out loud. A MCP server loading 664 tools into context is 300K to 650K input tokens before the agent has even read the task. A nine-tool allow-list is about 4,500 tokens. That's roughly a 100x cost cut plus least-privilege security from a single config field. He also notes that a casing mismatch once quietly switched off all tool filtering. Build your allow-list before your first production run. [READ MORE]

💀 The AI that confirmed the wrong price of toilet paper

Who Gives A Crap, the Australian toilet-paper company, sent a price-increase email, and one customer's copy had a typo: 48 rolls at $66 "becoming" 24 rolls at $69.50, more than doubling the per-roll price. The customer wrote in, alarmed. The AI-generated support reply cheerfully confirmed the wrong numbers back, "the quantity of rolls in your subscription will be halved, and the price will increase," because it had no way to check the email against the customer's actual subscription. The real price was $69.50 for 48 rolls. A company spokesperson confirmed the support reply email was AI-written; the company paused the agent and sent a correction. Give any customer-facing agent read access to the real record before it confirms anything. "The AI said so" means nothing if the AI only read the document you handed it. [READ MORE]

You don't send a colleague a three-word brief. You explain the context, the constraints, what you've already tried. But typing all that into ChatGPT takes forever — so you don't.

Wispr Flow lets you speak your prompts instead. Talk through your thinking naturally and get clean, paste-ready text. No filler words. No cleanup. Just detailed prompts that actually get you useful answers on the first try.

Millions of users worldwide. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

Alarm on absence, not error.

Will Kline runs a portfolio of agent-native software businesses from a single Windows workstation with one RTX 5090 in Oakland. The flagship is a social-posting SaaS for HiFi audio retailers, and it has a named first customer posting real content every day.

The setup reads clean. Nightly research agents monitor each client's sources, a composer drafts, a quality gate demands that every claim cite a source, and a publish-and-reconcile step posts to Facebook and Instagram, then checks the platform's actual state to confirm the post landed. An eight-lens audit swarm reviews the system, each lens blind to the others. A shared memory layer of about 450 files lets agents inherit past lessons. Every customer-facing post still needs a human tap to approve.

Then Kline let an AI audit swarm loose on his own code, and it surfaced the outages his dashboards had been hiding.

A background job once consumed a one-time OAuth claim 3.26 seconds before the customer finished the consent screen. The channel-bind failed silently, no error and no log, deterministic every time, and got blamed on popup blockers. For one billable account, nightly research fetches were blocked at cloud egress, so zero posts landed while the scheduler happily reported "cadence satisfied," because it measured whether the pipeline ran, not whether anything shipped. And a quality gate meant to enforce standards was grading full captions against an 84-character headline teaser, so it rejected every real draft and passed the empty ones.

Every expensive failure across all three was an absence. No channel bound, no post landed, no drafts produced, and every status light stayed green the whole time.

"My monitoring measured the things I had thought to measure," Kline writes, "which is a very different set from the things that break." And the line worth taping to your monitor: "A green dashboard over an empty feed is not a monitoring system, it's a lullaby."

Alarm on absence, not error. Instrument the terminal, real-world output, the thing that now exists in someone else's system, not the intermediate step that reports success. "The only number I fully believe anymore, for a billable account," Kline writes, "is landed greater than zero."

Two founders this week knew not to trust the word of an agent, and one company found out why they shouldn’t either.

Which agent in your stack are you trusting on its own say-so, and what real-world signal would tell you it actually did the work, not just that it reported it did?

Go find the signal that lives outside the agent's own report: the ledger, the platform, the customer's actual record. Wire your trust to that.

If this issue was useful, forward it to one founder who's running more on agents than they can currently see.

See you next Wednesday.
- Rich