Agentic AI
57 analyses from Devence Lab
· 2 min read
OpenAI's performance agents show why fixes need a second proof loop
The Performance Factory tracks over 1,700 investigations, but its strongest design choice is the post-deployment check. Agent workflows should close only when telemetry confirms recovery.

· 2 min read
XYBench shows agent evaluations must test the problem behind the request

· 2 min read
Kitesurf's open code makes agent browsing inspectable; access still needs a boundary

· 2 min read
PDD Kit makes the trust boundary the code review

· 2 min read
AgentCore live previews make port scope a release decision

· 2 min read
Google Cloud IAM Helper puts approval gates around agent access changes

· 3 min read
AWS's Wordle RL gains expose the gap between training reward and agent readiness
· 3 min read
OpenAI's Decisions API makes fast agent choices easier to test, not safe by default
· 2 min read
Agent Skills joining the AAIF changes governance, not the trust boundary
· 3 min read
NVIDIA's OpenShell moves agent authority beyond the runtime's control
· 3 min read
Smartsheet's MCP integration moves work-management permissions into agent execution
· 2 min read
LLM judges need their own audit before they can certify agent quality
· 3 min read
Cloudflare's bot trust bands turn crawler identity into ongoing accountability
· 2 min read
Taste-Bench exposes the wrong decisions inside long-running AI agents
· 2 min read
Tencent's unintended agent browsing is an egress-control failure

· 3 min read
AWS's Trust in Depth makes agent approval a policy-engineering problem
· 2 min read
Autonomous research agents need evidence review, not self-reported success
· 2 min read
Multi-agent teams must prove the cost of coordination
· 2 min read
Cloudflare's Deno acquisition makes runtime governance a platform dependency
· 2 min read
Goodfire's internal agent monitors reduce inspection cost, not accountability
· 2 min read
Gemini's enterprise agent needs an action boundary, not another chat policy
· 2 min read
722 AI maths manuscripts make verification the bottleneck, not generation
· 2 min read
Millions of agent requests are not just traffic. They are an identity-control failure.
· 2 min read
Self-improving agents need held-out gates, not just better benchmark scores

· 2 min read
Hard spend caps are becoming an agent safety boundary, not a billing feature

· 2 min read
Muse Gadgets turns hobby hardware into an agent authority boundary

· 2 min read
Claude Code Mods turn extensions into an in-process trust boundary

· 2 min read
Full Disk Access is becoming an agent privilege boundary, not a convenience toggle

· 2 min read
Always-on agents need standing authority limits, not one-time approvals

· 2 min read
Synthetic agent data needs executable verifiers, not plausible prompts

· 2 min read
Agent retrieval needs an exploit boundary, not just a browsing sandbox

· 2 min read
Clef turns agent routing into a decision-model problem. Set the policy before the model learns it.

· 2 min read
Agent sandboxes need communication boundaries, not just process isolation

· 2 min read
Model routing needs a quality floor. Cloudflare’s Auto Router makes cost policy executable.

· 2 min read
Always-on seller agents turn business rules into production code. Govern the workflow, not the prompt.

· 2 min read
Autopilot keeps working after the prompt ends. Govern the persistent identity, not the chat.

· 2 min read
69.4% became 45.9% when the server actually ran. Coding-agent gates need live serving.

· 2 min read
49% fewer tokens came from the harness, not the model. Benchmark the control layer.

· 2 min read
Agent observability should grade decisions, not just trace them.

· 2 min read
A central agent does not need centralised data. Keep authority in the owning account.

· 2 min read
An agent ignored 'no' and crossed the boundary. Scope needs runtime enforcement.

· 2 min read
A fluent agent can still choose the wrong skill. Evaluate routing and execution separately.

· 2 min read
444.6x cheaper is not the decision. Put probabilistic models behind typed boundaries.

· 3 min read
Real-time agents need two clocks. Conversation latency and task latency are different SLOs.

· 2 min read
2 seconds across 2 GB is not an agent latency result. Separate runtime start from model work.

· 2 min read
88% of AI proofs of concept do not need another model. They need a platform control plane.

· 2 min read
Gemini 3.8 Live does not end when the model stops speaking. Voice agents need a second state machine.
· 2 min read
Microsoft’s AI code is not a model policy. It is an application control requirement.
· 3 min read
OpenAI's RubyGems attack is not a containment failure. It is a disclosure failure.
· 2 min read
80% have embedded agents. 31% have deployed them. The gap is the whole story.
· 3 min read
Gartner's 40% is not a governance problem. It is a review-timing problem.
· 2 min read
When a thousand agents act as one, your identity model has already failed
· 2 min read
A $25M deepfake loss is an authorisation failure wearing a detection costume
· 2 min read
When the control is a human reviewer, the human is the attack surface
· 2 min read
A $5,000 query that every monitor approved
· 2 min read
Gartner's other warning: one governance policy across all agents causes the failure
· 2 min read
Early protocol decisions become systemic risk, and MCP is the current case study