Research
If you get early access to a model with safeguards turned down, you are now expected to run it in a hardened, monitored sandbox — the testing-side obligations are being written down.
Anthropic·2026-08-31
Research
Failure recovery and decomposition are where multi-step agents actually break, and this is a training recipe aimed at them rather than at a benchmark score.
arXiv·2026-08-24
Research
These are the authors’ own numbers on their own system and not yet independently replicated — but the claim is the interesting one either way: if most of a long-horizon failure rate is harness rather than model, the cheapest available win is in your own runtime.
arXiv·2026-08-15
Research
If you are picking an open model to build on, the report argues the leverage is in the ecosystem around it — derivatives, quantized builds, tooling — and that ecosystem is currently Qwen’s rather than the biggest model’s.
Hugging Face·2026-08-14
Research
If you fan out a fleet of identical agents, low variance is the failure mode: they make the same bet and fail together. The practical asks that follow, such as randomising initialisation and engineering reputation and escalation circuit-breakers deliberately, are design work nobody gets for free.
Anthropic·2026-08-13
Research✓ verified
Anyone running parallel agents on one repository is running this experiment already; the finding that identical parameters breed conformity argues for deliberately mixed models in a fleet, not one model cloned N times.
TechCrunch·2026-08-13
Research
This is a documented account of what an autonomous attacker actually does at machine speed across trust boundaries - the defensive reading for anyone about to hand an agent credentials.
Hugging Face·2026-07-28
Research
The reusable part is the shape of the work rather than the cipher: a multi-day, lightly-steered run produced a novel named technique, and several hundred hours of expert verification came afterwards — the verification is what made it trustworthy, not the autonomy.
Anthropic·2026-07-28
Research
Anthropic ran this research and is so far the only party to have reproduced it, so the numbers are its own finding until an outside lab checks the maths — but the background assumption should still move: AI-assisted cryptanalysis is demonstrated rather than hypothetical, which makes knowing your own cryptographic inventory the practical next step.
Anthropic·2026-07-28
Research
Prompt injection is still the open hole in every tool-using agent. A vendor claim of improvement is worth tracking, but until it is independently measured your permission gates stay exactly as load-bearing as they were.
Simon Willison·2026-07-25
Research
A persistent, growing skill library is the piece most agent stacks lack — every run starts cold — so the mechanism is worth tracking even before the numbers are public.
arXiv·2026-07-24
Research
First-hand, long-duration evidence for keeping deterministic gates and review layers in agent-driven development rather than trusting full autonomy.
HumanLayer·2026-07-23
Research✓ verified
Teams running agentic evals with relaxed refusals should audit sandbox egress first — the failure mode is ordinary infrastructure, not exotic AI.
Simon Willison / TechCrunch·2026-07-22
Research✓ verified
Loops, graphs, and anchors: how to orchestrate recurring agent work
If you run recurring agent jobs — nightly content builds, generated artifacts, scheduled reports — the practical lesson is to match the architecture to the job's real judgment need, not to reach for a framework. Most recurring work needs a fixed pipeline with one narrow place where an agent authors something, and the failure almost never lives in the loop-versus-graph choice — it lives in the seam: cleanup and 'did every output actually save' steps that were written as an instruction in the agent's prompt instead of as a hard step in the surrounding script.
AIU Research·2026-07-21
Research✓ verified
How AI Uni's agents remember: two kinds of memory, and why the durable one is layered
If you build agents that must remember anything across a killed session, a fresh subagent, or a model swap, the reusable lesson is that no single memory file can do the job: each layer here exists because a different, specific way work got lost was actually observed, and each delivers the fact at a different moment to a different audience. The honest-limits section is the most useful part — writing something down is not the same as guaranteeing it reaches the right agent later, and this design says so out loud.
AIU Research·2026-07-21
Research✓ verified
Cross-session agent memory in 2026: the platform now ships natively what most teams still hand-roll
If you're building an agent that needs to remember anything across sessions, check whether Anthropic's native memory tool already does what you were about to hand-roll — and if you're already building one, the field's clearest fix for the write-back/consolidation gap is a scheduled, importance-triggered pass, not more discipline.
AIU Research·2026-07-12
Research
Don't hang a pass/fail — or an accusation — on an AI-code detector. Treat its output as a hint to look closer, never as the verdict.
ccodelearner.com·2026-07-11
Research
There's finally a cheap way to prove a tool ran and what it returned — but only when you run the tool. A receipt a user hands you for a tool they ran themselves proves nothing.
arXiv (Basu, 2026)·2026-07-11
Research✓ verified
Can you prove an AI agent actually did the work? What today's tools can and can't verify
If your team accepts work an AI helped produce, this is the honest map of what's actually checkable in 2026 — and where a human still has to be the judge.
AIU Research·2026-07-11
Research
Anything an agent reports about work done on a machine you don't control is unverified by default. If it matters, re-run the check yourself instead of trusting the summary.
nhimg.org·2026-07-11
Research
Coding agents doing real repo work is exactly AI Uni's build model — a reminder that autonomous edits need guardrails measured on outcomes, not refusals.
arXiv (SABER)·2026-05-31