Abdullah Ahmad

[Abdullah Ahmad @abd__ahmad]

@dsp_ deserved — stateless is the first rev i can hand an ops team without the whole session-affinity conversation. that alone sells the upgrade

[Abdullah Ahmad @abd__ahmad]

@steipete real-ui testing is underrated — a good chunk of what we catch in browser-agent runs only ever shows up at the rendered-ui layer, never in api-level tests. video feedback loop is the unlock

[Abdullah Ahmad @abd__ahmad]

@ericzakariasson bridge pattern is the right call over maintaining n sdk ports. curious whether the wire protocol is the stable contract here, or if the adapter layer is what you'd steer people to build against

[Abdullah Ahmad @abd__ahmad]

@dexhorthy @josephruscio measuring outcomes is where every client conversation gets stuck for me — everyone wants one number and "tasks shipped" lies. did you two land anywhere on what's actually worth tracking?

[Abdullah Ahmad @abd__ahmad]

@HamelHusain skills pulling an entire docs site into context is the new left-pad. progressive disclosure was supposed to be the whole point of the skill format lol

[Abdullah Ahmad @abd__ahmad]

Most 'agent failures' I debug turn out to be tool design failures. The model did exactly what the tool description said — the description was just wrong. Write your tool docs like the model is a new hire on day one, because it is. Every single session.

[Cloudflare @Cloudflare]

Introducing Cloudflare Wallets. They will allow you to store stablecoins, purchase services, and receive funds across the web. https://t.co/ucAaoKMs4Q

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Abdullah Ahmad @abd__ahmad]

@trq212 the 80% cut tracks with what we saw in prod — most system-prompt bloat is compensating for a weaker model, and it turns into dead weight or outright conflicting instructions once the model can just infer intent. trimming helped our consistency way more than adding ever did

[Abdullah Ahmad @abd__ahmad]

@dsp_ the security side is where the exponential gets scary though — capability outran the isolation and auth story by a mile. running agents against real systems, the gap between "model can do it" and "safe to let it near prod" is where most of our eng time actually goes

[Abdullah Ahmad @abd__ahmad]

@_philschmid deprecating temp/top_p hits eval reproducibility harder than it looks — a lot of us pin temperature=0 for stable runs when regression-testing agents. what's the recommended path for deterministic-ish outputs now, or is that just gone?

[Abdullah Ahmad @abd__ahmad]

@HamelHusain @sh_reya the clustering step earns its keep for a reason people skip past — it surfaces the failure modes you never thought to write evals for. most of our prod eval suite came from clusters we didn't predict, not from the original spec

[Abdullah Ahmad @abd__ahmad]

@dexhorthy curious what slopcodebench actually penalizes — is it catching the "compiles + passes the happy path but invents an abstraction nobody asked for" case? that's the slop that survives review and bites weeks later

[Abdullah Ahmad @abd__ahmad]

The best agentic coding setups I've seen aren't the ones with the most tools. They're the ones where the agent notices it's stuck and stops guessing. Knowing when to ask is a capability, not a limitation.

[Abdullah Ahmad @abd__ahmad]

Welcome Opus 5

[Thariq @trq212]

Opus 5 rounds out our Claude 5 family beautifully.

I think it’s an incredible daily driver, pair it with Fable for planning, brainstorming or fixing the hardest bugs.

[Abdullah Ahmad @abd__ahmad]

The hardest part of agentic coding isn't getting the model to write code. It's giving it a fast feedback loop. An agent with tests, a linter, and a way to actually run the thing will out-ship one that just generates and hopes.

[Abdullah Ahmad @abd__ahmad]

The most underrated agent capability isn't reasoning — it's knowing when to stop and ask. A model that confidently plows ahead is a worse teammate than one that says "I'm not sure which file you meant."

[Abdullah Ahmad @abd__ahmad]

Half of prompt engineering now is deleting instructions, not adding them. Every rule you bolt on to fix one bad output quietly degrades ten good ones. The best system prompts I've seen this year got shorter, not longer.

[Abdullah Ahmad @abd__ahmad]

The hardest part of agentic coding isn't getting the agent to write code — it's building the taste to catch when the code it wrote is quietly wrong. Review is the new bottleneck, not authoring.

[Abdullah Ahmad @abd__ahmad]

The hard part of agentic coding was never getting the model to write code. It's writing the check that tells you the code is right. Skip that and you've just automated the part of the job you were already good at.