The LLM Context Window Is a Ceiling, Not a Budget. An Agent's Credit Card Works the Same Way.
Two posts this week, one arithmetic problem: a million token context window that serves 32K, and an agent whose real feature is the wall around its spending.
The week in one glance
- Kimi K3's open weights landed with a 1,048,576 token context window on the model card and a 32,768 token limit on its own reference serving recipe, both published July 27.
- Cerevisor 1.9.0 shipped two features that spend real money without a human pressing Run, and the part that took longest to build was the wall around them.
- Two posts is the whole week, and they turn out to be the same problem in two currencies: the number on the box is not the number that gets spent.
By the founder of Cerevisor
The theme: an LLM context window is a ceiling, and the budget is whatever survives
Two posts ran this week, one on local models and one on the harness, and I did not expect them to rhyme. They do. A model card advertising 1,048,576 tokens and a serving recipe capping the same model at 32,768 describe a ceiling and a budget, and the gap between them is filled by things nobody put in the spec sheet: cache, activations, concurrent requests, and the scaffolding a multi-agent harness wraps around every task. An unattended agent works the same way. The headline is that it can act on its own. The number that actually governs the run is the one that stops it. In both cases the interesting engineering is subtraction, and in both cases the number worth quoting is the small one.
What we published
AI coding agents this week
Cerevisor 1.9.0 ships Progressive Mode and Operator, both of which make paid API calls on a real provider key with nobody watching, so the write-up is about the trust layer underneath: a per-step ceiling of $0.50, a $5.00 per conversation budget, a daily wall that must be set before Operator can be switched on at all, a review queue where nothing applies without an explicit Accept, and an agent that disarms itself after three failed attempts.
Running models locally this week
Moonshot AI published Kimi K3's open weights on July 27 with a 1,048,576 token context window on the model card, and the release-day hardware guide listed the reference recipe as eight B300s with a 32,768 token limit, so the post walks through the four layers that spend a context budget before the task gets a turn: the serving config, the KV cache, the harness, and concurrency.
Signals to implications: context window math and AI agent guardrails
Signal. Same model, same day, 32 times apart. The 32,768 token reference config is not a consumer hardware compromise; it sits on 2,304GB of aggregate HBM.
Implication. Size the workload against the served limit, not the model card. If a capacity plan was built off an advertised window, it was built off the wrong number. [Eng Leader]
Source: Kimi K3 ships a 1M LLM context window. Your agent gets 32,768.
Signal. Ollama does not error when input exceeds the window. It truncates from the front, silently, and the OpenAI-compatible route ignores num_ctx passed through extra_body.
Implication. An agent that seems to forget its own instructions mid-run is often a serving config problem, not a prompt problem. The front of the prompt is the system instructions, so that is what disappears first. Fix it on the server, and check the local provider setup guide before rewriting the prompt. [Eng Leader]
Source: Kimi K3 ships a 1M LLM context window. Your agent gets 32,768.
Signal. Operator refuses to switch on until a daily budget exists, and when the wall is reached, paid checks stop until tomorrow with one notification rather than a dismissible warning.
Implication. When evaluating any unattended agent, the useful question is not what starts it but what stops it, and whether the stop is a setting the agent can talk its way past. The Operator guide spells out where each limit sits. [Founder]
Source: The AI agent guardrails I had to build before letting one spend money unattended
Signal. Nothing Operator drafts is applied automatically. Every proposal lands in a review queue with the evidence it was based on, and waits for Accept or Discard.
Implication. An agent that quietly installs its own improvements demos better and ships worse. Treat the review queue as a feature to look for in a vendor demo, not an inconvenience the vendor apologises for. [Exec]
Source: The AI agent guardrails I had to build before letting one spend money unattended
The contrarian take: the small number on the spec sheet is the honest one
The agent market still sells the largest number on the box, and the largest number is the least informative part of the specification. A million token window that serves at 32,768 tells a buyer almost nothing, while the four layers described in this week's context budget breakdown tell them everything about what the box will actually do under load. The same inversion runs through the unattended spend piece: nobody buys an agent that acts alone, they buy the confidence that it cannot ruin a Tuesday. Next time a capability number is quoted, ask for the served figure and the stopping condition. Those two answers price the thing.
Next week
Publishing cadence is deliberately light this month, so expect a small number of posts done properly rather than a daily stream. If this recap was useful, the newsletter delivers it straight to your inbox every Monday. Subscribe here.