buildOne report1 publisher MongoDB's 16 MiB document cap stalls agents that keep the whole transcript in one record
MongoDB rejects the entire write once a document passes 16 MiB, a hard cap that stalls agents keeping every turn in one snapshot. Frameworks can swallow that rejection, so the defence is measuring documents with $bsonSize and moving transcripts out early.
Reality
- Evidence60
- Adoption
- Not scored
- Hype gap0
- Incentives
- Not scored
- Confidence55
buildConfirmed2 publishers GHSA-864f-rcv7-6rh4 lets guest code hand the C++ bindings one type on the first read and a different one on the second. Fixes are in 7.0.1 and 6.2.0.
Publishers:dev.to · endorlabs.com Reality
- Evidence60
- Adoption50
- Hype gap+20
- Incentives60
- Confidence58
buildOne report1 publisher A five-minute function ceiling pushes long agent loops off Vercel's Hobby tier
Fluid compute caps Hobby functions at five minutes and Pro at 800 seconds, so an agent working a long task list needs a queue or a different host. The cheap VPS escape hatches have repriced too.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives65
- Confidence40
Arcjet puts a policy check in front of every action an AI agent takes
Arcjet's new runtime security takes agent activity in through the OpenTelemetry pipelines platform teams already run, then evaluates each tool call against Open Policy Agent rules before it executes and again after.
Reality
- Evidence32
- Adoption28
- Hype gap+30
- Incentives80
- Confidence52
buildOne report1 publisher A 2026 context-engineering guide pins agent failure on decisions that live outside the code
Durable state and sandboxed execution now arrive as platform primitives, so a dev.to guide argues the remaining failure mode is undocumented organizational context. It leans on one arXiv paper and one example.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+30
- Incentives45
- Confidence40
buildOne report1 publisher Salesforce puts a control plane above the harnesses that actually run its agents
Salesforce put Data 360, Informatica, MuleSoft, Tableau, Agentforce and Guardian under one new control plane on Thursday. The loops that run agents inside its own tools still belong to Mastra and the Claude Agent SDK.
Reality
- Evidence42
- Adoption15
- Hype gap+40
- Incentives75
- Confidence55
buildOne report1 publisher Checkpoint the step, not the pipeline: four agents and one unguarded embed
A lease-analysis pipeline built on Mastra gates three of its four processing stages behind skip flags read from the database. The fourth is chained unconditionally and re-runs on every resume.
Reality
- Evidence62
- Adoption12
- Hype gap+32
- Incentives60
- Confidence47
The arrayref compromise turned cargo update into the delivery channel
Wiz says a compile-time payload reached builds through a typosquatted dependency, and the attacker yanked every clean arrayref release so the responsible fix resolved to the poisoned one.
Reality
- Evidence66
- Adoption74
- Hype gap+12
- Incentives62
- Confidence58
A manifest edit, not a code edit: North Korea backdoored three Rust crates via typosquat
Wiz says the crates.io compromise executed at build time, so reviewing the library source would have caught nothing. Cargo did the rest.
Reality
- Evidence46
- Adoption71
- Hype gap+22
- Incentives68
- Confidence48
The sandbox teams fled vm2 for now has its own guest-to-host escape
Endor Labs says a type confusion in isolated-vm's ExternalCopy turns a single ivm.Reference into host control-flow hijack. The remedy is a version bump to 7.0.1 or 6.2.0.
Reality
- Evidence58
- Adoption66
- Hype gap+14
- Incentives66
- Confidence57
Spark's $22M bet that the agent framework layer can stay independent
Mastra raised $22M led by Spark Capital and launched a hosted platform in the same post. The framework was the cheap part; the runtime is where platform vendors already live.
Publishers:mastra.ai
Reality
- Evidence20
- Adoption34
- Hype gap+42
- Incentives90
- Confidence28