For a while, Tokenmaxxing was a thing. Leaderboards, how many tokens can you use. It seamed a bit crazy, but hey, BigTech is gonna BigTech. Until they didn't.

At Tend, we are very much NOT big tech. Healthcare in NZ is a fairly low-margin industry, after the clinicians have been paid, so we can't just blow big wads of cash on AI tooling, any more than we can hire 5 more engineers.

On the engineering team, we pretty much exclusively use Amp as our agent harness. It's proven to be stable, reliable, and generally just really good. Like Linear (which we also love), it has taste and an opinion, and that fits to what we are trying to do. I'm coming to the conclusion that having founders - or senior leaders - who are not just American (Finnish, German, Australian) is a good thing. (no shade on Quinn tho).

Amp, however, is (or was) very much a usage-based tool. They didn't mark up the tokens, but you paid for your usage. We were dropping $300-500 per day for a bit, and it escalated quickly with a major project rewrite from mobile app to web.

So this resulted in a peak months usage of around $6k USD. If you translate that back to NZD, thats basically a fully loaded, well paid senior/staff engineer. It's not going to break the business (I hope!) but if we can optimise it down without losing velocity or developer happiness, then we should.

On the plus side, it's not 1999/2005 and we are not paying $2500/year/engineer for Visual Studio and MSDN!

So how did we do it?

This all started around April or so. OpenAI had the best (or at least close to the best) models, and unlike Anthropic, allowed people to use their subscriptions with 3rd parties (this was the time of OpenClaw going nuts).

Amp integrated this into their product, so we could link our OpenAI/ChatGPT accounts, and our spend drops massively.

We got each engineer to open a ChatGPT account (and expense it) - one even found a 2-accounts-for-the-price-of-one offer. Usually its the base $20/month USD one, and if they were hitting limits more than once a day, upgrade to the $100 plan (if not just use tokens from Amp).

Don't ask - just upgrade and expense it. It's still going to be cheaper.

This changed our spend from close on $6k/month (USD) to around $500. At worst it'd be capped at $800 USD. Thats fairly easy to justify given the performance/productivity improvement.

At around the same time, Amp introduced Orbs and Puck. We've moved a lot of our development to them, so people really don't need the $4000 USD mega laptops that we have (M5 Pro / 48GB / 1TB). Sadly we can't get a refund on those and move people to M5 Airs...

It's not a dictate to move to use them. It's just easier.

Someone asked on X:

This is it for us:

• Get an OpenAI account per engineer and wire it up to Amp (not a company one, more than 100 people and you pay token prices!). This saved us SO much money.
• Don't fret about "is this the best model this week". The actual gap between GPT6 and Opus5.5 is minimal. Don't bother with Fable/Astra except to experiment. You don't need it. medium is excellent, high if you actually think you need it. TRUST THE HARNESS.
  

• Get at least a Amp Megawatt plan for orbs and just stop thinking about the cost - just use them (a1.large min size! No need to skimp). I've never even come close to the limits - and you get the $20 in token credit! Hobby is fine if you are just playing with it, or for people like Product Managers or Designers who also want to use Amp a bit.
• Consider using some of the open weights models - K3, GLM - they are excellent and cheap AF. Especially for code review.
• Get Amp (ie, Puck) to add the plumbing to make Orbs work for your repo. Sometimes, you can't (our backend needs to talk to AWS/RDS and thats a bit much to put in there, tho it's possible), but mostly... you can. PR previews are dead to us, we use the Orb-hosted Portal.
• Embrace the Mac client. My flow is now (if I'm not using an Orb, which I mostly do) :
  • Fire up the TUI in a folder with amp --no-tui
  • Fire up the mac client and use that folder/instance.
• Once you've tried it a bit, turn on "amp.dangerouslyAllowAll": true, - you can control what it can/can't run, but so far I've never had it do something bad. Our AWS stuff is aws-vault gated tho - YMMV. One shot a full iOS app FTW.
• Don't manage your context. Just talk to the agent. The harness manages it for you, if it needs to. Subagents are your friend, too - tell the main thread (or Puck) to use them. I get good mileage out of a high level controller which spawns medium level subagents. All in Orbs.

One of the best flows I have right now:

This is a skill (full text is longer but not a lot):

Given a Honeycomb trace and the honeycomb MCP, investigate the problem, look at the code and see if you can definitivly work out whats wrong. If so make an Linear ticket (via MCP) with context, and make a PR I can review

Then I find a honeycomb trace on Slack and tell Amp to investigate it

And land it as needed. It's not 100% - sometimes the fix isn't obvious. But I've resolved a lot of noise in our ops channel from this one trick.

I tend to do them over coffee in the morning, and then when I'm at my desk, I have 4 or so PR's to look at and land (or not).