Almost nothing published today is about a model getting better. It is about the parts that surround one: what stops an agent that decides wrongly, what the context window costs when nobody is accounting for it, and what replaces the diff once a human has stopped reading it. Two conference talks and a survey paper arrive at that claim from different directions, which is that reliability now lives in the harness. The releases follow the same line, with Cursor, LangChain and Nous all shipping plumbing rather than intelligence.
Over 30% of changes now merge with no review at all, and Ankit Jain's answer is not to restore the diff but to replace it: capture the agent session as acceptance criteria, build a registry from your own recurring review comments, then verify against a live preview. A survey of coding agent reliability lands on the same target from the research side, arguing the defect teams blame on the model usually sits in the system operating it.
Theo has lived in a terminal since he was thirteen and now argues it is the wrong surface for coding with agents. Set next to the review argument, the shape is the same: an interface built for a person typing commands is being asked to host a process nobody is watching line by line.
A dated brief from the vibe-coding frontier. Today’s Intel.
Context work produced the sharpest number of the day. Retaining reasoning across turns together with compaction moved one model from 13.3% to 38.3% on the same task, a change most agent loops can make without touching the model. A conference session puts a cost model and a recall number against compaction, and a measurement study asks the unglamorous version of the question: does a language server actually save a coding agent tokens over grep.
The releases are all plumbing. Cursor put out Origin, its own code hosting platform, in beta and synced from GitHub. LangChain and AWS shipped middleware that lets a tool call settle a 402 against a session budget and retry, with the payment recorded next to the reasoning that caused it. Nous turned Hermes agent profiles into named bots that each keep their own role, model, memory and skills.
Two independent takes on containing an agent that acts wrongly. Ryan Dahl presented a security firewall for agents, and a separate breakdown splits the problem into infrastructure, runtime and network layers, holding credentials outside the agent and inspecting outbound calls before they leave the boundary. Both start from the assumption that the agent eventually makes the wrong call, and design for what happens next.
Three papers press on the same soft spot in agent evaluation. Hand-written judging rubrics over-credit, so one induces the rubric from environment reward instead. Another scores the damage a half-finished attempt leaves behind rather than only the end state. The third makes the case that benchmarks quietly measure replacement when the system being built is collaborative.
Cost accounting got three sources at once: a routing paper showing that per-token price understates real spend because hard tasks inflate token counts by tier, a year of production serving traces on caching and load balancing, and a hands-on look at how honestly marketplace providers actually cache what they claim to.