Your agent can return a perfect 200 OK and still do the wrong thing.
It can choose the wrong tool, hand stale context to a subagent, loop through retries and burn money while every normal dashboard stays green.
That is why this week’s biggest shift is not another smarter model. It is finally getting a useful window into how agents behave after we deploy them.
🔥 The Big One
Cloudflare is turning agent behaviour into something you can debug
Cloudflare introduced a dedicated Agents view and agent-aware tracing on 4 August.
The dashboard connects the parts normal application monitoring misses: model calls, tool execution, approvals, subagents, token use and the Workers infrastructure underneath them. You can inspect a trace as a waterfall or replay the recorded context from a session, including the messages, reasoning, tool arguments and results that shaped the final answer.
That replay is recorded data, not a re-execution. This matters because the useful question is rarely “did the API respond?” It is “why did the agent believe this was the right action?”
“An agent can return HTTP 200 and still fail.” (Cloudflare)
My take: observability is becoming part of the agent product, not an optional DevOps layer.
If a normal app fails, you inspect logs and traces. If an agent quietly picks the wrong tool, the failure can look like a confident, valid response. Builders need to see which model ran, what context it received, where approval paused the turn and which tool call pushed the result off course.
Cloudflare’s tracing currently supports OpenTelemetry-compatible integrations including Think, Flue and AI SDK. Traces can also be exported to compatible destinations. Payload recording can be switched off when messages or tool results contain secrets or personal information, which should be the default question before anyone records full agent sessions.
Cloudflare is framing this as the start of a wider agent development lifecycle. The useful part today is narrower and real: deployed agent behaviour is becoming inspectable.
That changes the shipping checklist. Do not stop at “the agent completed the task.” Ask whether you can explain the path it took.
Read Cloudflare’s launch post →
⚡ What shipped this week
1. IntelliJ’s Java and Kotlin brain is coming to Cursor and VS Code
It supports mixed Java and Kotlin projects, completion, navigation, analysis, refactoring, plus Maven, Gradle and Bazel. The preview is free, new builds arrive every two weeks, and each build has a 30-day evaluation period. JetBrains says an IntelliJ IDEA Ultimate subscription will be required after the preview.
“IntelliJ IDEA’s Java and Kotlin intelligence is available in a preview extension for third-party editors, starting with VS Code and its forks (such as Cursor).” (JetBrains)
This is bigger than editor preference. Coding agents get better when the environment can answer deterministic language questions instead of spending tokens guessing from text. JetBrains is also testing the same LSP capability with terminal agents, but it has not published those results yet.
2. VS Code now shows what every subagent is doing
VS Code 1.131 makes parallel agent work easier to supervise.
The Agents window now shows each running subagent’s model, elapsed time and active tool call without forcing you to open every conversation. Its progressively rolling-out Agent Host also runs Copilot, Claude and Codex agent systems in dedicated processes through the Agent Host Protocol.
“Because a session lives in its own process, the same session can be connected to and rendered from multiple VS Code windows at once.” (VS Code)
The practical win is attention management. Parallel agents only help if you can see which one is progressing, which one is stuck and where your judgment is needed. Agent Host is still opt-in through chat.agentHost.enabled, so treat it as an evolving architecture rather than a finished default.
3. Vercel gave agents private homes inside one Sandbox
Vercel Sandbox now supports multiple Linux users and groups.
Each agent can run as its own user with a private home directory. Agents cannot list, read or write each other’s private files. When collaboration is required, a group can open an explicitly shared workspace.
“Each agent runs as its own user with a private home directory.” (Vercel)
This is the right default for multi-agent builds: isolate first, then share only the directory the task requires. “They are all inside one sandbox” should never quietly mean “they can all read everything.” Normal Vercel Sandbox access and usage pricing still apply; no separate price was announced for this capability.
4. Prelint checks whether AI wrote the right product
Prelint was Product Hunt’s number one launch on 29 July with 481 votes and 114 comments. Its official site positions the product as a review layer that checks pull requests against product specifications, not just whether the code compiles.
It is aimed at the failure agents are especially good at hiding: clean code that ignores an architectural decision, invents a requirement or builds the wrong workflow.
“AI doesn’t slow down when it’s wrong.” (Prelint)
The catch is obvious and important. A reviewer grounded in stale specs can enforce yesterday’s decision perfectly. The winning workflow is not “let another AI approve the PR.” It is keeping product intent current, surfacing conflicts and leaving the final decision reviewable by a human.
🧰 Worth your time
-
GitHub Copilot comment-triggered automations: Run exact automations from issue or pull-request comments for docs, error investigation and follow-up work.
-
Cursor’s Google Workspace plugins: Agents can read and write across Gmail, Drive and Calendar. Useful, but sending mail and changing calendars deserve narrow scopes and review-first defaults.
-
Medley missions for Claude Code and Codex: Turn a larger outcome into a supervised task graph with parallel workers. The public plugin is inspectable, while its downloaded mission engine is proprietary, so review that trust boundary before installation.
Which tool did it choose? Where did it retry? What context reached the subagent? Could you explain the final action to another developer?
What is the one agent decision you wish your current tools made easier to inspect?
Hit reply and tell me.
I read every single one.
Talk soon PAPAFAM,
Sonny 👋🏼
👇🏽 Don't forget to follow me across socials!
Responses