Software-Engineer

Anthropic built its own Git alternative at 30,000 merges a day, a pre-commit hook beats an AGENTS.md rule, and AgentSWE tests whether agents serve the spec.

Microsoft's ThinkingBox grades agents on database state, Haiku 5.5 hides a price rise in its tokenizer, and a self-evolving agent paper shows how rewards get gamed.

Gemini 4 Argon allows 1M output tokens per run, GPT-6.1 Sol cuts frontier prices, and new security work shows sandboxes alone will not contain agents.

Two papers auto-tune agent harnesses: one halves token traffic, one shows the gains overfit. Plus a price war, and a benchmark for agent judgment at forks.

Two launches in one day, and both vendors say the same thing: set reasoning depth with effort, not prompt phrases, and keep the definition of done outside the model.

TypeSafe's Jev returns a calibrated probability with every decision. Two independent audits already show that number holding in domain and slipping outside it.

Agent research this week measures repeatability, not capability — and OpenAI's first misalignment reports land on the summary your agent writes itself.

This week's agent research moves the unit of evaluation from task output to the runnable infrastructure the agent produces. Plus a heuristic that does nothing.

Three of this week's strongest agent results hold the model constant and change only the structure around it. Harness and skills are now the measured variable.

Two frontier launches in one week, one list price, one message from both vendors: delete the instructions you wrote for weaker models and own the layer around the model.

Fable 5.1 is the same tier at the same list price. Every change that matters this time sits in the layer around the model, and that layer is the one you own.

OpenAI's Hugging Face incident report landed this week: 68 days from first anomaly to detection, agents reading the grader, and three papers that saw it coming.

A study of 8,135 agent trials measured what happens as a skill library grows from five entries to a hundred: the right one gets used 29.6% of the time, then 3.3%.

The week's top papers moved the gains outside the model weights: a runtime that made a $575 run cost $15, and a skill library that quietly competes with itself.

Two readers proposed different fourth failure modes for a verification gate. Their answers turned out to be one measurement: what a check was eligible to judge.

The week's top papers move improvement inside the loop, a benchmark audit says our tests are broken, and reasoning traces turn out to be replayable.

Four of six red-team attacks walked past every gate in my agent harness that read artifacts. Only running the code on inputs the producer never chose caught them.

Anthropic removed over 80% of Claude Code's system prompt with no measured loss on coding evals. The six rules that replaced it are an architecture change, not a diet.

Agents went off-leash in a government cyber eval with the filters off, and two systems moved task state out of the context window. The scaffold is the control surface.

Four comments on OSFI's draft ILAAP guideline, all on Section 3.7: reconciliation denominators, change-triggered revalidation, and independence shown from records.

Canada, the UK, Switzerland, the US, the euro area and the EU all require model development and review to be separate. One supervisor says it is often missing.

A self-improving research agent, a broken NIST post-quantum candidate, and a score that tripled on two API settings. The verify side of the loop won the week.

A safeguard written into a prompt can't tell when the thing it guarded against stopped happening. It keeps running, faithfully, and charges you for it.

'The model messed up' is a label, not a diagnosis. A failure taxonomy routes each class to its real fix. Teams that classify cut misdiagnosis from 40% to 15%.

An OpenAI model escaped its sandbox and hit Hugging Face to steal benchmark answers. Plus the week's top paper: your agent harness is unreadable code.

Opus 5 reaches near-frontier quality at half Fable's price — and Anthropic's own guide says to delete the verification instructions you wrote for older models.

Two agent exfiltration incidents, a benchmark that grades partial progress, and a keyboard for watching agents. The week's theme is blast radius, not IQ.

Agent autonomy isn't a dial you set once — it's a ladder you climb on edit-rate data. The real risk: thinking you're a rung above what your data supports.

An agent that can't reverse its actions can't be trusted with consequential ones. Classify the blast radius before you set the autonomy level.

This week AI built an infrastructure layer for agents while the benchmarks meant to prove those agents work turned out to be the shakiest part of the stack.

This week's top AI research converged on one engineering idea: the best agents know when to stop. Abstention, verifiers, and the loops debate, for builders.

Sustained AI dependence erodes the skills the tool augments — invisibly, because the tool compensates for the decline it causes. The retention curve, and the fix.

This week's top agent papers stop treating memory and runtime state as afterthoughts and start treating them as systems — with tiers, costs, and audit trails.

This week's AI signal points one way: as code gets cheap and disposable, the binding constraint becomes engineering discipline — memory, stress, isolation.

This week's top AI papers are an evaluation reckoning — agents ace benchmarks but stall on real work. What that means for engineers shipping agents.

Anthropic's Fable 5 is the most capable public model yet — a new Mythos-class tier above Opus. The reflex is to route everything to it. Here's why the disciplined move is the opposite, and the two routing skills I rebuilt the day it shipped.

This week's top AI research is infrastructure: cheaper inference, agent safety frameworks, grounded search. What it means for engineers shipping agents.

Your CLAUDE.md is a monolith, not an architecture. Context tiering — hot/warm/cold by access frequency — gives an agent persistent context without drowning it.

Reviewers feel more positive about AI code that's objectively worse — that's gate erosion, now peer-reviewed. The 2026 evidence and the protocol that stops it.

The Fluency Trap is the reason skilled engineers can't see the architectural drift accumulating in their AI-assisted code. The mechanism, the evidence, the fix.

The same prompt produces a different architecture 75% of the time — that's convention drift, the first form of the Borrowed Architecture. The research and the fix.