Agents · industry

Agents in Software

Agent deployments and launches in Software, from the Drops.

From the Drops

AgentsSoftwareTestingHuman in loop: unclear

Preprint: a logging-style MCP injection succeeded in 62.64-100% of attacks

In a preprint on Log-To-Leak, researchers report poisoned MCP tool descriptions made agents send queries and responses to an attacker-controlled call after legitimate work, across 18 model-server pairs.

Why it matters: Finishing the task is no sign of safety: 16 pairs changed task completion by at most 6.63 points. The authors call for host-controlled authorization of secondary tool calls.

arXiv · Oct 28

From Drop #051 · Wed, Oct 28 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: coding agents found under 21% of the bugs maintainers fixed across 20 repos

PROBE covers 2,457 bugs fixed over 18 months. The authors report recall below 21% and precision below 9% against that set, yet LLM and human evaluators judged 85% of findings valid.

Why it matters: Agents find real bugs, just not the ones maintainers fix. Stronger models raised recall from 8% to 17% but not precision, the authors report, so triage still falls on people.

arXiv · Oct 28

From Drop #051 · Wed, Oct 28 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: 725 of 1,403 agent errors self-recovered, often within a few steps

Researchers built ReInBench from 1,259 agent traces on computer-use and web tasks. They report self-recovery differs by model, and their ReCoF predictor improves Macro-F1 by 9.02% over the strongest baseline.

Why it matters: Stepping in on every error wastes human attention. The authors suggest intervening on errors unlikely to self-recover and letting the agent continue otherwise.

arXiv · Oct 28

From Drop #051 · Wed, Oct 28 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: 31.1% of agent runs succeeded while failing a harness reliability check

TraceTrial-Bench scores agent harnesses on seven components, from context to verification. Across 4 harnesses, 9 models and 95 tasks (3,420 runs), the authors report harness scores from 59.8% to 72.0%.

Why it matters: A passing task can hide a broken harness: the authors report 31.1% of runs succeeded despite failing at least one harness criterion.

Hugging Face Papers · Oct 23

From Drop #050 · Tue, Oct 27 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: many agent memory systems underperform plain RAG over raw history

Researchers evaluated memory systems across multiple benchmarks and model backbones and report many do worse than naive retrieval from raw history. They call the losses from reorganizing or distilling facts a memory tax.

Why it matters: Before adding a memory layer, compare it with simple retrieval: the authors say memory processing can remove or obscure the evidence an agent needs.

On the Frontier watch →

arXiv · Oct 27

From Drop #050 · Tue, Oct 27 · Evening

AgentsSoftwareLiveHuman in loop: unclear

Preprint: 37.4% of agent apps that must keep state lose it when the container restarts

Across 433 deployed open-source agent systems, researchers checked 14 contracts from LangGraph and platform docs. Of those needing saved state, 41.5% lost it on container replacement; only two in three fixes fully worked.

Why it matters: If your agents pause for human input or resume saved work, test restarts and container swaps before production. The authors found 78% of fixes changed app code.

On the Frontier watch →

arXiv · Oct 27

From Drop #049 · Tue, Oct 27 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: the same tools via MCP or a CLI changed agent behavior in 94.3% of episodes

Across six models and three environments, researchers report matched interfaces gave different scores and costs. Their DART repair raised scores by up to 9.0 points and cut MCP token cost by up to 27.1k tokens per task.

Why it matters: MCP versus CLI isn't a neutral choice. Test both on your own tasks before standardizing on one.

On the Frontier watch →

arXiv · Oct 27

From Drop #049 · Tue, Oct 27 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: prompt-caching bugs in agent code mostly fail silently; 18 fixes followed

Studying 120 caching pull requests across 10 open-source agent systems, researchers built CACHELINT. It flagged 32 mis-integrations across 571 projects at 0.86 precision, leading to 18 upstream fixes.

Why it matters: A broken cache costs money without throwing errors. Check your cache hit counts against what the provider bills.

arXiv · Oct 27

From Drop #049 · Tue, Oct 27 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: realistic MCP tool faults cut agent task scores by 60%, and scale didn't help

Researchers built MCP-Injector, with faults calibrated from 57,240 threads across 1,774 MCP repos. Task-completion score fell from 8.08 to 3.23 out of 10; a 120B gpt-oss kept no more of its score than a 20B.

Why it matters: The authors report agent stacks differed in absorbing faults, suggesting framework error handling matters more than model scale. Capability benchmarks can hide this.

On the Frontier watch →

arXiv · Oct 26

From Drop #047 · Mon, Oct 26 · Morning

AgentsSoftwareLiveHuman in loop: yes

LangChain shares the five parts of its internal company agent harness

LangChain says its sales, engineering and marketing agents live in Slack, keep global and per-user memory, run code in sandboxes, and often run on schedules or events. Email tools that act for a user work only in DMs.

Why it matters: LangChain calls identity the thorniest problem: some tools act for the user, others use fixed credentials, and some actions are limited to approved users.

On the Frontier watch →

LangChain · Oct 25

From Drop #047 · Mon, Oct 26 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: context compaction broke 5-64% of 22 coding-agent successes

Compact-Bench pairs full-history and summary-only runs across 14 summarizer models. LLM summaries kept at most 16.0% of recurring details, versus 79.1% for mechanical extraction, the authors report.

Why it matters: Long-running coding agents lean on summaries to keep going. The authors say retaining relevant information alone does not ensure the agent can finish the task.

On the Frontier watch →

arXiv · Oct 26

From Drop #047 · Mon, Oct 26 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: tool output that omits limits misled agents and the verifiers checking them

In a production RAG system, correct tools printed output omitting search limits or result caps, the author reports. In one test, stating those details flipped verifiers from accepting wrong answers 90 of 90 times to rejecting them 90 of 90.

Why it matters: The author says an audit prompt run with Codex and Claude Code found such formatters in OpenClaw and Hermes; the OpenClaw fix has been merged.

arXiv · Oct 26

From Drop #047 · Mon, Oct 26 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: plain RAG over raw history beats complex agent memory in 6 of 7 settings

Researchers report Raw-RAG, dense retrieval plus reranking over raw history, ranked first in six of seven benchmark-model settings across four memory benchmarks, beating the strongest agentic memory system by 0.13-5.53 points.

Why it matters: The authors say preserving raw information matters and call naive RAG an essential baseline for the conversation tasks they tested.

On the Frontier watch →

Hugging Face Papers · Oct 22

From Drop #047 · Mon, Oct 26 · Morning

AgentsSoftwareTestingHuman in loop: no

Preprint: agents taught while they practice write skills that transfer better

Researchers report Socratic Skill Learning, where an Educator guides a Student agent that writes its own Skill, reached 0.801 utility on its main 155-case cohort, versus 0.723 with no Skill and 0.761 for Trace2Skill.

Why it matters: A focused variant hit 0.805 with 61.7% fewer Educator tokens, the authors say: a hint at cheaper ways to build reusable agent skills.

Hugging Face Papers · Oct 13

From Drop #045 · Sun, Oct 25 · Morning

AgentsSoftwareTestingHuman in loop: no

Preprint: IF-MAP picks which agent in a multi-agent team to improve first

Researchers report that across six multi-agent workflows, tuning the prompts of the agents IF-MAP selects gained 3.81 percentage points on average, versus 1.85 when agents were picked uniformly.

Why it matters: The authors say it cut Qwen3-8B token use by 74.2% versus optimizing every agent: a cheaper way to aim improvement effort in agent teams.

On the Frontier watch →

Hugging Face Papers · Oct 17

From Drop #045 · Sun, Oct 25 · Morning

AgentsSoftwareTestingHuman in loop: no

Preprint: judge agent memory by whether it improves future actions, not recall

Researchers propose Memory-Centric Continual Learning: a taxonomy of memory substrates, assets and policies, a test of retention under interference, adaptation and transfer, and a framework for how memory scales.

Why it matters: Most agent memory is scored as a retrieval cache, the authors argue; their protocol aims to separate real behavior gains from simply adding context.

On the Frontier watch →

Hugging Face Papers · Oct 21

From Drop #045 · Sun, Oct 25 · Morning

AgentsSoftwareTestingHuman in loop: no

Preprint: a world model flags hijacked GUI agent screens at 0.987 AUC in 17 ms

Researchers report in a preprint that LGWM, trained on 1.85M real GUI transitions, predicts the next screen's representation after an action and flags mismatches, such as a login screen shown after a tap on View order.

Why it matters: A hijacked login screen can look legitimate. The authors say judging the transition, not the screen, gives computer-use agents a cheap check, tested on their own benchmark.

arXiv via Hugging Face · Oct 8

From Drop #044 · Sat, Oct 24 · Evening

AgentsSoftwareTestingHuman in loop: no

Preprint: DeLoop cuts repeat-action loops in long-running agents from 57.3% to 6.1%

Researchers report agents often re-issue identical tool calls, more so once history is compressed. Training on their own loops cut Qwen3-8B tasks with loops from 57.3% to 6.1% on CodeGym, using about a third as many steps.

Why it matters: Loops waste tokens in long runs. The method needs no task rewards or LLM judges, though the abstract reports results only on Qwen models; on Qwen3-32B success rose from 37.7% to 44.8%.

On the Frontier watch →

arXiv via Hugging Face · Oct 14

From Drop #044 · Sat, Oct 24 · Evening

AgentsSoftwareLiveHuman in loop: unclear

LangChain: retrieve, offload and delegate to keep an agent's context clean

LangChain says its Deep Agents harness saves tool results over about 20,000 tokens to files and hands the agent a path and preview, and suggests subagents read retrieved files so the main agent sees only conclusions.

Why it matters: Search results can flood a context window. LangChain also warns retrieved text can carry prompt injection that no prompt or delimiter strategy fully prevents.

LangChain · Oct 23

From Drop #044 · Sat, Oct 24 · Evening

AgentsSoftwareLiveHuman in loop: yes

JetBrains' open-source ThinkRail gives each coding agent its own git worktree

ThinkRail, an Apache-2.0 JetBrains InnovationHub project, gives each agent workspace its own git worktree and branch, and keeps project intent in a graph of spec files the agent reads. 47 points and 72 comments on HN.

Why it matters: Running parallel agents without touching the main branch is a common pain; this is one open design to try or borrow from.

ThinkRail · JetBrains Hacker News · discussion

From Drop #044 · Sat, Oct 24 · Evening

AgentsSoftwareLiveHuman in loop: unclear

LangChain details how Deep Agents offloads and summarizes context to fight context rot

LangChain says its open-source harness writes tool results over roughly 20,000 tokens to files, summarizes older messages at 85% of the context window, and keeps long-term memory in a plain file you control.

Why it matters: The ideas carry to any harness: offload rather than truncate, keep compaction recoverable, and own your memory so you can switch models.

On the Frontier watch →

LangChain · Oct 24

From Drop #043 · Sat, Oct 24 · Morning

AgentsSoftwareLiveHuman in loop: unclear

Databricks: Astra rollout raised AI coding spend about 60%; routing cut costs up to 90%

Databricks says it rolled Astra out to about 3,500 engineers and spend rose about 60% without clear gains on everyday tasks. Moving to cheaper models, routing and spend visibility cut some unit costs by as much as 90%.

Why it matters: Databricks says token price is a poor predictor of per-task cost, and the same model in a different harness cost more than 2x as much. Measure cost per task, through one gateway.

Databricks · Oct 22 Databricks · Oct 22

From Drop #042 · Fri, Oct 23 · Evening

AgentsSoftwareTestingHuman in loop: yes

DoorDash agents fixed 45 of 50 stale feature flags at $4.79 each, engineers report

In a preprint, DoorDash engineers describe a multi-agent system that reads live flag data over MCP, processes Jira tickets and opens pull requests. Engineers approved removals for 45 (90%) of 50 stale flags, averaging 13.8 minutes each.

Why it matters: DoorDash counts 1,659 stale flags across 623 repositories, at one to two hours per manual cleanup. The test covered 50 flags: a narrow, well-defined task agents handled cheaply.

arXiv · Oct 23

From Drop #042 · Fri, Oct 23 · Evening

AgentsSoftwareLiveHuman in loop: yes

NVIDIA's OpenShell agent sandbox reaches 0.1 and draws 451 points on Hacker News

OpenShell's repo says it confines each agent with kernel-level policy on files, system calls and network, adds credentials only to requests for approved endpoints, and uses formal verification to hold risky policy changes for human review.

Why it matters: It is an open-source, Apache-2.0 way to give agents real tools without handing them secrets. NVIDIA says Jensen Huang described it in Astana as a way to contain AI agents.

On the Frontier watch →

GitHub · NVIDIA/OpenShell Hacker News · Oct 22 NVIDIA · Oct 23

From Drop #042 · Fri, Oct 23 · Evening

AgentsSoftwareAnnouncedHuman in loop: unclear

Preprint compares MCP and CLI for agent tools and recommends a hybrid

Across 14 evaluation dimensions, the authors find CLI more efficient for single-user, local workflows, while MCP covers needs CLI cannot: fine-grained access control, credential isolation, structured audit and multi-tenant governance.

Why it matters: The authors argue progressive disclosure brings MCP's token use close to CLI's, so production teams need not trade governance for context budget.

On the Frontier watch →

arXiv · Oct 23

From Drop #042 · Fri, Oct 23 · Evening

AgentsSoftwareLiveHuman in loop: unclear

Anthropic gives Agent Skills to the Agentic AI Foundation, now home to 10 projects

The Linux Foundation says AAIF added Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba, joining MCP and A2A. It says AAIF now has 305 members.

Why it matters: Skills, a cross-provider API spec, RL environments and sandboxes now sit under neutral governance, so builders can adopt them with less single-vendor risk.

On the Frontier watch →

Linux Foundation · Oct 22 Linux Foundation · Oct 22 Simon Willison · Oct 22 X · @ClaudeDevs · Oct 22

From Drop #041 · Fri, Oct 23 · Morning

AgentsSoftwareLiveHuman in loop: yes

Rescale's test-triage agent now agrees with human triage 68% of the time

AWS says Rescale built a Claude Opus 4.8 agent on Bedrock that classifies nightly test failures and drafts fixes for engineer review. Agreement with human triage rose from 44% to 68%, and time blocking deploys fell 20%.

Why it matters: A concrete pattern: the agent does the first pass, engineers approve before any action, and an evaluation harness has tracked accuracy since day one.

AWS · Oct 22

From Drop #041 · Fri, Oct 23 · Morning

AgentsSoftwareTestingHuman in loop: unclear

LangSmith LLM Gateway puts spend caps, rate limits and fallbacks in front of agents

LangChain says the gateway, in public beta for Plus and Enterprise plans, enforces spend caps at four levels, rate limits, model fallbacks and PII redaction across providers. A capped agent gets a 402 error.

Why it matters: LangChain's example: an agent stuck in a retry loop overnight makes 10,000 calls. Central caps stop that without rewriting every agent's code.

LangChain · Oct 22

From Drop #040 · Thu, Oct 22 · Evening

AgentsSoftwareLiveHuman in loop: yes

Latent Space: a Claude Code side agent flagged risky uncommitted work

A Latent Space author says Claude Mods, plugins that reshape Claude Code, proved useful in weeks of use: a built-in side agent flagged that the edited file and about 45 others held uncommitted work.

Why it matters: It's a concrete supervision pattern: a second agent watches the first and surfaces only what a human should see. The author notes Anthropic ships six built-in mods.

On the Frontier watch →

Latent Space · Oct 21

From Drop #040 · Thu, Oct 22 · Evening

AgentsSoftwareTestingHuman in loop: no

Preprint: agent-written verifiers often reject valid solutions

Researchers report in a preprint that across 19 models, normalized balanced accuracy on VerifierBench's 222 tasks ranged from 10.6 to 62.1. Most false rejections came from verifiers that accepted no tested solution.

Why it matters: If agents check their own work or filter training data with generated tests, the authors suggest checking that a verifier accepts correct solutions, not just that it rejects bad ones.

arXiv · Oct 22

From Drop #040 · Thu, Oct 22 · Evening

AgentsSoftwareTestingHuman in loop: yes

Preprint: PerMIT gates agent memory so untrusted entries need confirmation

Researchers report PerMIT gates untrusted, environment-derived memory, with integrity theorems machine-checked in Lean 4. They report zero observed attack success on Trojan Hippo and shared knowledge across up to 90 agents.

Why it matters: The authors call persistent memory a largely undefended surface for injection. Reducing attacks to one auditable confirmation is a design teams can study.

On the Frontier watch →

arXiv · Oct 22

From Drop #040 · Thu, Oct 22 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: all 6 frontier LLMs tested fell for subagents reaching beyond their scope

Researchers' SUA-Arena has 6 attack strategies and 188 test cases. They report all six frontier LLMs were vulnerable; GPT-5.6-Sol as supervisor hit an attack success rate of up to 0.93.

Why it matters: The authors report prompting and restricted communication did not reduce the risk consistently, and say multi-agent systems need explicit ways to enforce authority boundaries.

On the Frontier watch →

arXiv · Oct 21

From Drop #038 · Wed, Oct 21 · Evening

AgentsSoftwareLiveHuman in loop: yes

Linux Foundation survey: GenAI cuts code authoring time 52%, but 25% of PRs are slop

LF Research says respondents estimate a 52% cut in code authoring time, while maintainers estimate 25% of incoming PRs are low-quality AI slop, 31% in large projects. Only 35% of projects have a formal GenAI policy.

Why it matters: For teams using or maintaining open source, LF points to written AI policies, disclosure and testing; 21% of respondents say complex tasks take longer with GenAI.

Linux Foundation · Oct 20

From Drop #036 · Tue, Oct 20 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: a tool gate blocked every protected action agents wrongly requested

Researchers report that on invalid-proof cases Nemotron and Llama requested protected actions in 52.1% and 54.2% of trials, yet a deterministic tool gate committed none. A valid proof from a compromised signer still passed.

Why it matters: Check what agents execute, not only what they ask for: enforcement at the tool held, but the authors note proof validity is not factual truth.

On the Frontier watch →

arXiv · Oct 20

From Drop #036 · Tue, Oct 20 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Cohere: a strong lead with cheap agent workers neared frontier results at 70% lower cost

On 200 ProgramBench tasks, Cohere says a Sol lead with 16 cheap Luna planning workers hit 61.8% reward vs 63.2% for an all-Sol team, cutting cost from $510 to $155. Teams gained up to ~10 points over solo agents at 3–7× the cost.

Why it matters: Cohere found where you put agents matters more than how many: delegating only planning matched delegating every stage (63.6% vs 63.2%) at half the cost.

On the Frontier watch →

Cohere · Oct 14

From Drop #034 · Mon, Oct 19 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: MS-Eval finds sub-agent delegation often raises cost with uneven payoff

Testing four main-agent and sub-agent designs with eight frontier models on MCP-heavy tasks, researchers report specialized sub-agents work best. Delegation helped most open-source models but cost closed-source ones performance.

Why it matters: Gains came from context offloading and real parallelism, the authors say; unnecessary or redundant delegation turned the same cost into pure overhead.

On the Frontier watch →

arXiv · Oct 19

From Drop #033 · Mon, Oct 19 · Morning

AgentsSoftwareTestingHuman in loop: unclear

GatewayBench preprint tests seven MCP gateways on 524 tasks with 11,118 real tools

Researchers built tasks across six domains from 251 MCP servers, injecting irrelevant tools to mimic large deployments. They report substantial gaps between gateway designs in task completion, efficiency and resilience to attacks.

Why it matters: The authors say gateway effects interact strongly with the model and harness, so the middleware between agents and tools deserves its own testing.

On the Frontier watch →

arXiv · Oct 19

From Drop #033 · Mon, Oct 19 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: CarryOn cuts long-horizon agent tokens by 51.9% on OSWorld 2.0

CarryOn organizes an agent's context around reusable task objects instead of compressed histories. The authors report gains of 5.8% on OSWorld 2.0 and 7.0% on BrowseComp-Plus, with token use down 51.9% and 22.9%.

Why it matters: Long agent runs keep re-reading how they got somewhere; this preprint tests carrying forward the progress itself, a design question for any multi-hour harness.

On the Frontier watch →

Hugging Face Papers · Oct 13

From Drop #033 · Mon, Oct 19 · Morning

AgentsSoftwareLiveHuman in loop: yes

Anthropic: staff's agents logged 6 times their hours by August

Anthropic says its employees' agents worked 6 times as many hours as their human handlers by August. Following 192 adopters against 74 non-adopting engineers, code added rose sharply; agent code was no more often reverted.

Why it matters: Anthropic says human attention is now the bottleneck: reading agents' output would take half a non-technical employee's work week. It notes adoption wasn't random, which may bias estimates.

Anthropic · Oct 17 Hacker News · Oct 18

From Drop #032 · Sun, Oct 18 · Evening

AgentsSoftwareLiveHuman in loop: unclear

DeepSeek opens Search and Fetch APIs for agents at $0.75 per 1000 requests

DeepSeek's docs say Search returns structured web results from its in-house engine and Fetch returns a page as Markdown. Both are billed per successful request; failed calls, empty results and timeouts are not billed.

Why it matters: DeepSeek says tool calls made when its API runs via Claude Code, Codex or DeepSeek Harness already hit these APIs free; per-request billing starts November 1, 2026, Beijing time.

DeepSeek API Docs · Oct 18 Hacker News · Oct 18

From Drop #032 · Sun, Oct 18 · Evening

AgentsSoftwareAnnouncedHuman in loop: unclear

Google: Gemini agents pause at hard spend caps; coworker agents get their own accounts

Google says coworker agents get their own Workspace account, and a hard cap set in Cloud Billing automatically pauses a project's agent. It says these Gemini agent capabilities are not publicly available yet.

Why it matters: Runaway bills and audits slow agent rollouts; Google pitches spend caps, a cryptographic Agent Identity, sandboxing and prompt injection defenses as answers for startups' security reviews.

Google Cloud · Oct 15

From Drop #032 · Sun, Oct 18 · Evening

AgentsSoftwareLiveHuman in loop: yes

Sphinx, a Git hook, quizzes developers on agent-written code before it ships

Shown on Hacker News (193 points, 61 comments), Sphinx pauses pushes on protected branches and asks what the diff does when things go wrong. Its makers say skips are allowed but logged.

Why it matters: It checks that the person shipping agent code understands it. The makers say answers never reach the coding agent and .env files and keys are never opened.

Sphinx · Oct 18 Hacker News · Oct 18

From Drop #032 · Sun, Oct 18 · Evening

AgentsSoftwareLiveHuman in loop: unclear

LangChain ships LangSmith Engine v2 red teaming and Managed Deep Agents v0.9

LangChain's October newsletter lists Engine v2, which red-teams agents before production, plus trace search and fine-tuning from production trajectories. Its evals guide advises checking LLM judges against human scores.

Why it matters: Agent tooling is shifting from building agents to testing them. LangChain says failures found in production should become new eval examples.

LangChain · Oct 15 LangChain evals guide · Oct 15

From Drop #031 · Sun, Oct 18 · Morning

AgentsSoftwareAnnouncedHuman in loop: unclear

Preprint formalizes self-evolving agent harnesses that revise prompts, tools and code

Researchers describe agents that revise their own scaffolding while model weights stay fixed, with accepted changes persisting across runs, and give bounds for admitting revisions using fresh validation data.

Why it matters: The authors say reliable deployment needs continual validation and independent safeguards for learning, runtime privileges and consequential actions.

arXiv 2610.18813 · Oct 13

From Drop #031 · Sun, Oct 18 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Preprint: frontier agents fail over 58% of closing actions in multi-agent tasks

Researchers built 600 computer-use tasks where sub-agents report outdated or conflicting info. Testing 15 frontier models, they found weaker ones follow misleading peers and fixing conflicts is far harder than spotting them.

Why it matters: The authors say even frontier models often forget to release shared locks or hand off tasks, a check worth adding if you run teams of agents.

arXiv via Hugging Face · Oct 13

From Drop #030 · Sat, Oct 17 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: a local small model with selective cloud calls cuts agent cost up to 62%

Researchers report EdgeWise runs a small model on-device and hands key steps to a cloud LLM, matching or beating pure-LLM runs while cutting inference cost up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1.

Why it matters: The authors find not every step of a long agent run needs a frontier model, which points to a way to trim token bills.

arXiv via Hugging Face · Oct 13

From Drop #030 · Sat, Oct 17 · Evening

AgentsSoftwareTestingHuman in loop: unclear

Preprint: MAI-Gym's 9B GUI agent reports 76.2% on OSWorld-Verified

Researchers release MAI-Gym, an open platform for GUI agent training. Their 9B agent scores 78.4% on MobileWorld, and RL in a mocked train-ticketing app lifted success on a real phone from 28.8% to 48.1%.

Why it matters: The authors release the framework, 2,000 verified tasks and the checkpoint, so teams can test whether training on mock apps carries over to real ones.

arXiv via Hugging Face · Oct 14

From Drop #030 · Sat, Oct 17 · Evening

AgentsSoftwareLiveHuman in loop: yes

Anthropic: Claude Code merged PRs more than doubled as the revert share fell over 40%

Anthropic says Claude writes about 90% of Claude Code's code. Claude verifies its work, every PR gets AI and security review, a classifier sets the risk tier, and an agent's approval never counts as a human's.

Why it matters: A concrete playbook for agent-volume code: Anthropic says PRs with substantive review comments rose from 16% to 54%, and it caps re-review at three rounds after one PR looped over fifty.

Claude blog · Oct 16 Hacker News · Oct 16 @ClaudeDevs on X · Oct 16

From Drop #029 · Sat, Oct 17 · Morning

AgentsSoftwareTestingHuman in loop: unclear

Cohere: a Sol lead with cheap workers hit 61.8% at $155 vs $510 for an all-Sol team

On 200 ProgramBench tasks, Cohere says a Sol lead with 16 cheap Luna planning workers reached 61.8% reward vs 63.2% for an all-Sol team. Delegating only planning matched full delegation at half the cost.

Why it matters: Agent teams beat solo agents by up to about 10 points but cost 3 to 7 times more, Cohere says. Where you put the workers matters more than how many you add.

Cohere · Oct 14

From Drop #029 · Sat, Oct 17 · Morning

AgentsSoftwareLiveHuman in loop: unclear

Stack Overflow post: MCP made tools easy; running them securely is the gap

A Hexum.dev builder on the Stack Overflow blog says teams keep rebuilding auth, retries and logging around MCP, and cites a 2026 study: 40.55% of 7,973 live remote MCP servers exposed tools without authentication.

Why it matters: If agents call your tools, the author's advice is least privilege, explicit approval for destructive actions and treating tool results as untrusted input.

Stack Overflow Blog · Oct 15

From Drop #028 · Fri, Oct 16 · Evening

Showing the newest 50 of 60. Older ones stay on their Drop's page in the archive.