daniel

understanding the world around me

[daniel @dansap__]

@dani_avila7 Might need to recreate this on a small scale. Great read 👍

[daniel @dansap__]

@Damilolawo1a People charging your paid account 💸

[daniel @dansap__]

@ashuuu_soft AWS but mostly to learn these systems more in depth

[daniel @dansap__]

@JoshARosen @warpdotdev Now it’s just a matter of ensuring that the agent knows what to evaluate and doesn’t just give itself a pat on the back 😅

[daniel @dansap__]

@GergelyOrosz I could see them being used like some of the last gen models were on an enterprise level. Being used by companies to look through an entire codebase for security issues, etc.

[daniel @dansap__]

@shivangitech_97 Feeling the urge to say Go 😅

[daniel @dansap__]

The amount of times my agent told me this and then when pressed, did a full 180

[Hugo @hugospdx]

I think “task completed successfully” should require evidence.

Tests passed.
Diff reviewed.
Requirements matched.
No unrelated changes.

Otherwise it’s just the agent congratulating itself.

[daniel @dansap__]

How should new SWE balance breadth with depth when growing their knowledge?

[daniel @dansap__]

@sydneyrunkle Isn’t the main point of a subagent to allow the new agent to perform its best so you give it only the most necessary parts of the current context?

[daniel @dansap__]

@lawofcons Do you think that leetcode/DSA is still worth it for people looking to get into big tech?

[daniel @dansap__]

@thesamparr I’m a founder and have $3 revenue, can I join?

[daniel @dansap__]

@ashoKumar89 What is the best way to avoid n+1 queries ?

[daniel @dansap__]

@LeoBuilds_ knowing how to build guardrails + deterministic checks to make sure the agent stays on track is huge too

[Santiago @svpino]

I haven't read a single line of AI-generated code in a couple of months.

Manually reading code is no longer effective for me.

Code is a low-bandwidth information source. The more you have, the harder it becomes to separate the signal from the noise.

Instead, I spend most of my time designing processes to automatically verify the code, write unit tests, evaluate the system while it runs, handle potential errors, and log and trace everything that happens.

Here is what the ideal cycle looks like:

Plan → Code → Deploy → Test → Monitor → Learn

The goal is to run this cycle automatically, end to end. The interesting part of this comes in the final three steps:

1. Test: We stress-test the running system by generating fake traffic or diverting some of the live traffic to it.

2. Monitor: Observability is always baked in as part of our coding practices, so here we monitor what happens with the system as much as we can.

3. Learn: We scan the telemetry data from the monitoring process and feed any errors into the plan for the next iteration to fix those issues.

@Dynatrace is one way to implement this pattern. You can use it to connect your agents and retrieve runtime information to feed into the planning process.

Basically:

• It will give you the telemetry data you need
• It will let your agents retrieve the runtime signals

For automated agents, you can run this feedback loop as a hook. Every change then passes through deployment, testing, observation, and analysis before completion.

Here is an explanation of the complete workflow:

https://t.co/QIQr20jYKA

Thanks to the Dynatrace team for partnering with me on this post.

[daniel @dansap__]

need this 💀

[daniel @dansap__]

For anyone that’s been able to use Astra, how much of a difference has the model made in your day to day work ?

[daniel @dansap__]

Is Cursor worth trying over Codex/Claude desktop app + VSCode?

[daniel @dansap__]

@devXritesh How do you scale up from this smaller scope for agents? So far skills and using an orchestrator agent has been helpful

[daniel @dansap__]

@techyoutbe harness and loop engineering?

[daniel @dansap__]

@oldstackjournal Doing this especially when starting on a new code base