Suraj Suradkar

Building Xeyria 🚀
Building in public.
Turning ideas into intelligent execution.

[Suraj Suradkar @SurajSuradkarfo]

@DeepPatternAI Hi, I saw your email and have completed the beta application. I’ve also sent my response by email with the requested details. I’ll make time for the testing and feedback once the next steps are shared. Looking forward to it.

[Suraj Suradkar @SurajSuradkarfo]

@williamm168 @suidevelopers @reginaldojsf @b1ackd0g @bbarwik The shift from “we got an audit” to “how do we know testing is actually complete?” feels important. Continuous verification loops seem like the harder problem now.

[Suraj Suradkar @SurajSuradkarfo]

@Suryansh_2001 The speed is real, but the debugging tax later can be brutal. I wonder if an “understand the architecture before execution” step could prevent a lot of this.

[Suraj Suradkar @SurajSuradkarfo]

@ivanburazin This is the part that still feels under-discussed. Even with good model guardrails, broad local access can create a huge blast radius. Curious if you’ve seen any practical patterns for making agent permissions both scoped and auditable?

[Suraj Suradkar @SurajSuradkarfo]

@DeepPatternAI Sure. You can reach me at [email protected] and I’ll share the details there. Happy to continue the testing conversation privately.

[Suraj Suradkar @SurajSuradkarfo]

@DeepPatternAI Got it. I’d be happy to test it in a real coding-agent workflow and share detailed feedback, especially around the quality/evidence checks.

[Suraj Suradkar @SurajSuradkarfo]

@Dayhaysoos Makes sense. I'd watch for one gap: AST can verify the shape, but not whether the underlying assumption is actually true. A cheap repo-state check alongside it might catch that.

[Suraj Suradkar @SurajSuradkarfo]

@jalalash I like the idea, but I'd measure more than token reduction: what gets evicted and whether later steps need it. Otherwise, smaller context may look better while increasing recovery and rework.

[Suraj Suradkar @SurajSuradkarfo]

@_VaibhavBajpai The hard part isn't running multiple agents—it's keeping their state clear: who changed what, why, and what's unresolved. A lightweight state record per agent may reduce review overhead more than another orchestration layer.

[Suraj Suradkar @SurajSuradkarfo]

@shegtory This feels like the next bottleneck for coding agents: generating options is becoming cheap, but evaluating them well is still hard. Taste, context, and tradeoff awareness may become the real developer advantage.

[Suraj Suradkar @SurajSuradkarfo]

@DAIEvolutionHub Interesting approach. The Git-based context is especially interesting because it makes project knowledge persistent without relying on the agent's session memory. I wonder how it handles context that becomes outdated after architectural changes.

[Suraj Suradkar @SurajSuradkarfo]

@_alejandroao The interesting part is that persistent memory isn't just storage. Retrieval, relevance, and deciding what context is worth keeping are equally important. That's where agent memory gets really challenging.

[Suraj Suradkar @SurajSuradkarfo]

@jun_song I’ve noticed this too. Maybe part of the cost is agents keeping more project context and tool state alive. As agents become more persistent, RAM could become an architecture problem, not just a hardware problem.

[Suraj Suradkar @SurajSuradkarfo]

@DeepPatternAI Sure, happy to continue privately. You can DM me here and I’ll share my email/contact details.

[Suraj Suradkar @SurajSuradkarfo]

@DanKornas Staleness indicators are a really important piece. Storing a decision is easy; knowing when its underlying assumptions changed is much harder. How are you detecting that today?

[Suraj Suradkar @SurajSuradkarfo]

@vhsdev @lydiahallie This is underrated. Without the rejected alternatives and their reasoning, an agent can resurrect old ideas as “new” solutions. How are you capturing those decisions today?

[Suraj Suradkar @SurajSuradkarfo]

@ai_magg The “tried and abandoned” part is huge. Most memory systems preserve decisions but lose why alternatives were rejected. How are you keeping that history in sync with the codebase?

[Suraj Suradkar @SurajSuradkarfo]

@bogdan_veliscu The re-explanation cost gets painful fast on larger codebases. How are you preserving the reasoning from one session to the next, not just the raw context?

[Suraj Suradkar @SurajSuradkarfo]

@GreenlandMa1 CLAUDE.md growing into 200+ lines is such an interesting pattern. At some point, how do you keep old rules from becoming stale or conflicting with newer ones?