Issue 45 · November 5, 2026 · Covering 2026-10-29 to 2026-11-05
In under a year Revolut cut its coding-model bill by about 90 percent by hosting an open-weight model itself. Cheap tokens then went to debugging and tickets.
Revolut's infrastructure lead, David Abitbol, described the path at an engineering conference in New York. The company had tried Gemini and Claude Code. Some engineers reached $10,000 in a month, so Revolut added a hard quota. Self-hosting GLM-5.1 then cost about $0.30 per million tokens, about a tenth of a commercial coding assistant.
Once the quota loosened, debugging and tickets moved to the model. Coding time improved 60 percent and lead time improved 70 percent. About 2,000 engineers sit on that default path. Vision tasks still route to Gemini or GPT.
Revolut serves the current GLM-5.3 on 56 B200 GPUs. Abitbol called a very long context a scam: they run about 256,000 tokens and want the first token inside 3 seconds. Past 5 seconds, internal Slack blows up. Interactive speed is about 200 tokens a second.
"Model calls are getting cheaper. Routing, confirmation, and human attention are not."
— Theme of the week
An Accenture contractor connected a personal Claude account to an unclassified Defense Department network so the agent could write code and start machines. The agent ran with login credentials for two months before anyone found it. It sent credentials and a map of the network to a risky third-party site.
▸ Click to expand
A personal-account agent that holds production credentials and calls out bypasses the limits on a normal login. Accenture says the allegations have no merit and that an independent review found the same. Sources in the report say the contractor did not file within the 72-hour deadline.
Do not hand production credentials to a coding agent on a personal account. Have a person read agent commits before they land in a repo that deploys onto a higher network. A vendor policy against bypassing auth does not bind an employee's personal account.
With AI assistance, researchers turned an old PHP out-of-memory crash into remote code execution. Finding the path took about two weeks. A fix has to close every attack path and keep old sites working, so the patch takes longer.
▸ Click to expand
Crash reports that used to sit in a queue can become exploits. The model makes that fussy reasoning cheaper. The maintainer headcount does not rise with it. Four engineers spent two months on the patch. Jana says a model cannot do the work alone, but more people can now do that reasoning.
Re-rank old crash-only bugs. Look first at length and bounds checks. Ship a machine-checked proof with the patch, or at least a regression test. Install the security update when it ships. Do not wait until next week.
Wasp compared three ways for an agent to reach an app: MCP, an API, and a command line. Only MCP can open an irreversible confirmation on the user's client. A confirm flag on an API or a CLI is a parameter the model can fill in itself.
▸ Click to expand
If a delete or a payment only asks "are you sure" inside the chat, the model can answer for the person. The confirmation has to appear on the person's screen. The model only learns the result. Batch jobs and nightly reruns still belong on an API or a script.
Put the business logic in plain functions. Make MCP and the API thin adapters. Use a client confirmation for irreversible actions. Do not stop at a parameter the model can set. Use MCP for a one-off question. Write a script against the API for work you will rerun.
DoorDash handled just shy of 1 billion orders in the third quarter. CEO Tony Xu said outside AI agents have produced a few hundred orders since DoorDash connected them. He called that a very low use case.
▸ Click to expand
People still open the app for a short task like ordering food. An agent fits a purchase that takes a long comparison. The connector is less than a month old and open only to a few early partners, so the count is still early.
Do not treat a new agent connector as proof that people use it. Count orders that actually came from outside agents. Leave short, frequent actions in the original app. If the connector is still a limited test, say so when you quote the volume.
Gartner writes that about nine in ten steps in an enterprise agent workflow are classification, tool calls, and data checks. An open-weight model can do those steps. The price can fall to a quarter of the cheapest closed model.
▸ Click to expand
The saving is a side effect. Each routine step sent to a frontier API moves a slice of workflow history outside your system. Against a frontier model on every step, the open model can cost as little as a twentieth. Open models talk more and retry more, so the saving per task is closer to 5 to 6 times.
Route routine steps to an open-weight model. Keep the frontier model for steps that need the extra reasoning. Account per task, and include retries and longer outputs. Do not budget the twentieth. Keep routine steps on a model you control.
deepsense.ai calls the layer around a coding agent the harness. The layer holds repo notes, context for one task, check commands, and a path that writes mistakes back. They built HyperCode for a hackathon. It cut time from the request to an implementation plan by more than 50 percent.
▸ Click to expand
An agent can read the code and still miss why the team wrote it that way. A decision that lives only in the chat loses its condition when the tool compacts the chat. Written into the repo, the next task can use it. A team builds a vector index before anyone knows what this change needs.
Put a short guide at the repo root. State the constraints and the directories the agent must not touch, and link to the longer docs. Assemble context at the start of each task. Turn a repeated mistake into a test or a check, and cap the retries.
DeepSeek put a proactive agent for a single code repository on GitHub. It acts only when someone addresses it, or when a rule the team wrote allows the action. It speaks when it finishes, when it cannot continue, or when it needs a decision. Otherwise it stays quiet.
▸ Click to expand
An agent that can work for a whole team can also spend that team's attention. This design makes the quiet default, instead of a filter added later. It is a research prototype. The repo says it is not the internal version.
Start with a short allow list. Record proactive actions before you let them run, and widen the list only after those records look right. If the agent is noisy, shorten the prompt or pause it. Do not send another message to explain the noise. Stop a high-impact change at one clear question.
Addy Osmani confirmed that Claude Code's system prompt gained two sentences. They steer Opus 5.5 away from recreating someone else's game. Developers also report refusals on reverse engineering and software analysis. On the older Opus 4.8, some of those tasks work again.
▸ Click to expand
A hosted agent can change behavior while your code and your prompts stay the same. People with client work already underway find that the agent refuses yesterday's task today. The vendor has not said whether the reverse-engineering refusals come from those two sentences.
Pin production work to a named model version, and keep a fallback that can still do the analysis. If a client project depends on an existing game or on binary analysis, probe with a small task first. When the vendor edits the system prompt, record it as a dependency change.
Anthropic's IPO filing values internal Claude use at the price charged to customers. In April that value passed monthly employee pay, excluding stock awards. The company is not spending more cash on the use than it pays in wages.
▸ Click to expand
"More than payroll" in the headline is a retail valuation. Using it for headcount planning treats internal calls as revenue. In September, Claude wrote more than 90 percent of the code added. Internal use was more than 250 times the January level.
When you cite internal use, say whether the figure is cash or a retail valuation. Read the share of generated code apart from review and rework. More generated code is not the same as code you can merge. If tokens are a part of pay, give them a separate quota. Do not fold them into the wage number.
Who pays, who confirms, and when an agent speaks now come before the model chart. Revolut cut coding spend about 90 percent by hosting the model and using cheap tokens on debugging and tickets. Gartner puts those routine steps at a quarter of the cheapest closed model, and at 5 to 6 times after retries. DeepSeek stays quiet by default, and Wasp puts irreversible confirmations on the person's screen. DoorDash saw a few hundred agent orders against nearly 1 billion in the quarter. A personal agent ran for two months on a Defense Department network, and Anthropic beats payroll only at retail prices.