Tesla Just Put Your Agents on a $200 Allowance
The token-cost reckoning has arrived — and a new paper called SelfCompact suggests the fix isn't a spend cap, it's teaching your agents to shut up.
Your Agents Have a Hoarding Problem
Tesla just handed every employee a $200-per-week token allowance and told the ones who blow through it to go beg a manager for more. Read that as a memo about frugality and you've missed the story. This is the first honest admission from a Fortune 50 that agentic AI has a runaway-cost problem nobody scoped in the pilot. At Kuaray, here's our take: the spend cap is a bandage on a wound the industry gave itself. The real disease is that your agents are context hoarders — and there's finally a paper that treats the illness instead of the symptom.
The Bill Nobody Modeled
For eighteen months the pitch was "just add tokens." Then the invoices landed. Uber reportedly torched its entire 2026 AI budget in four months. Now Tesla is metering usage by the week. This is what the industry started calling tokenmaxxing — agents that solve a task and also drag every dead-end, every stale tool call, and every abandoned chain-of-thought along for the ride, paying full price to re-read their own junk on every single turn.
Here's the part your FinOps deck doesn't say out loud: a long agent trajectory isn't just expensive, it's dumber. Stale context doesn't sit there quietly. It anchors the next generation, pulling the model back toward derivations it already abandoned. You're paying premium rates to make your agent worse.
SelfCompact: Let the Agent Take Out Its Own Trash
Enter SelfCompact (arXiv:2606.23525), out of a Johns Hopkins-led group. The idea is almost insultingly simple, which is usually the sign it's right.
Today's scaffolds compact context on a fixed token threshold — hit 100k, summarize, keep going. That's a dumb trigger. It fires mid-derivation, guillotining a half-finished proof, or mid-search, torching results the agent was about to use.
SelfCompact hands the model a compaction tool and a one-page rubric for when to pull it:
- Fire when a sub-task just resolved, or the trajectory is clearly converging.
- Suppress when you're mid-derivation, or genuinely stuck and still need the breadcrumbs.
No fine-tuning. No reward model. Just inference-time judgment. The numbers: up to +18.1 points on math, +5–9 on agentic search over a no-summarization baseline — at 30–70% lower cost per question. Cheaper and smarter. That combination almost never shows up on the same slide.
What To Actually Do Monday
1. Instrument before you cap. A blanket $200 wall is management by blunt trauma. Trace your agents, find where tokens actually die, and you'll usually find hoarded context, not legitimate work.
2. Make compaction a decision, not a threshold. If your scaffold summarizes on a token count, it's throwing away partial results at random. Move the trigger to task structure.
3. Treat context as a managed resource. The teams winning in H2 2026 aren't the ones with the biggest token budget. They're the ones whose agents know what to forget.
Schedule a Technical Architecture Review with our Strategists — we help engineering leaders build agent systems that get cheaper as they get better, instead of the other way around.