Your company runs on one provider

This week a single government order took every operator outside the US offline overnight.

A US export-control order pulled Fable 5 for all foreign nationals. Operators in Germany woke up to a CEO that couldn't think.

Most operators build their company on one adapter, set the model once, and never think about it again. That's the part that bit hardest this week.

Here's the pattern 👇

---

The week two decisions took down half the org charts

First it was Fable 5, yanked by export control for every foreign national worldwide. Operators outside the US lost their main model with zero notice.

Then came the Agent SDK credit scare. "Are we on API consumption today, or is this the last day on subs?" The whole community spent days doing budget math on a billing change that was due to land any minute.

Then it didn't. Anthropic paused the credit change, claude_local stayed free, and everyone exhaled.

But look at what actually happened. For days, operators with a healthy 6-agent company couldn't tell you whether next week ran on their subscription or got billed at full API rates.

→ One decision lands somewhere above you.
→ Your CEO can't think.
→ Your workers stall mid-sprint.
→ You find out from a Reuters headline, not a dashboard.

That's not a model problem. That's an org-design problem.

✦	Lesson A company that runs on one provider isn't a company. It's that provider's outpost with extra steps.

A mixed-adapter team is not a luxury anymore

Here's the thing: Paperclip was built on adapters for exactly this week.

Claude Code, Codex, Gemini, OpenCode, Hermes, and a dozen more all sit behind the same agent interface. The CEO doesn't know or care which model is answering. It just gets answers.

Which means the question was never "which model is best." It's "what happens to my company when the best one disappears for a Tuesday."

A single-adapter company has one failure mode that takes everything down at once. A mixed-adapter company degrades instead of dying.

One provider down on a single-adapter team is a dead company.
One provider down on a mixed-adapter team is a slow afternoon.

You don't need to run four models. You need your company to survive losing one.

✦	Lesson Spreading your agents across adapters is the cheapest insurance you'll ever buy. You pay for it in 30 minutes of setup, once.

Match the model to the role, not to the budget

The other thing this week surfaced: most operators pick a model for the company when they should be picking one per role.

The budget-model thread was the most active of the week, and the operators who'd actually done the work all said the same thing.

Operators in the community peg Gemini Pro somewhere between Opus 4.7 and 4.8 on agentic work, at roughly a tenth the cost. Minimax M3 runs about $500 a year for 5.1 billion tokens a month. K2.7 is a solid coding model. Step 3.7 is the one people reach for when an email needs to sound human.

Cheap models are genuinely good at some jobs and genuinely bad at others. One operator ran market-intel scraping on a budget model and got great results, then tried to write sales emails with it and hit a wall. The content read like a robot.

That's not a reason to abandon the cheap model. It's a reason to stop using one model for the whole company.

→ Market intelligence and scraping → a cheap, capable model.
→ Code and refactors → a strong coding model.
→ Customer-facing writing → the model with the best voice.
→ Strategy and planning → your most capable model, run rarely.

★	Tip Your most expensive model should run on your lowest-frequency, highest-stakes role - the CEO setting strategy on a 4-hour heartbeat, not a worker firing every 15 minutes.

The CEO thinking once every four hours on a premium model costs you almost nothing. A worker grinding through a queue on that same model is where the bill explodes.

?

Quick ask show me your adapter setup. I'm collecting real Paperclip configs this week - which adapter each role runs on, model per agent, monthly spend, agent count. Reply or DM me yours and I'll send back the one role I'd move to a different model first. This is exactly how we figure out which defaults need to change 🙏

The 30-minute resilience pass

You don't need to rebuild anything. You need to stop running every agent on the same provider.

Open your company and look at the adapter on each agent. If it's claude_local all the way down - CEO, builder, reviewer, everyone - you have one failure mode for the whole org.

That's the gap. Not that you're on Claude. That every agent is.

!	Warning Don't wait for the next outage to find out your whole company sits on one provider. By then your sprint is already stalled and you're debugging under pressure.

Here's what I'd do this weekend.

Pick your two most critical roles - usually the CEO and your main builder. Put them on different adapters. If the CEO runs on Claude, move the builder to Codex or Gemini. Run a test issue through each to confirm the work still lands.

There's no automatic failover - Paperclip runs one adapter per agent, so when a provider goes dark you switch the affected agents by hand. But that switch takes seconds, and a company split across two providers never goes fully dark.

Then set a per-agent budget cap so the cheaper models can run hot without surprising you. It's a field on the agent itself - a monthly spending limit, where zero means no limit - separate from the heartbeat settings.

Pick the adapters. Set the caps. Then stop tweaking.

✦	Lesson Resilience isn't running more models. It's making sure no single provider can take your whole company down.

---

The useful question after this week isn't "which provider is cheapest" or "which model is smartest."

It's "what survives when my main one goes dark for a day."

A single-adapter company answers that with silence.

A mixed-adapter company answers it with a slightly slower Tuesday.

---

That's it for this week.

Until next week,

Keep delegating

Aron 📎

---

P.S. If you only harden one agent this weekend, {{firstName}}, make it the CEO - a company whose CEO can't think doesn't degrade, it stops.