Harel Asaf

AI Builder @Elementor | Designing multi-model systems for product & ops | Ex-corporate lawyer who now builds the stack | 🎙️ Podcast host | Tel Aviv

[Harel Asaf @RL_Asaf]

@n8n_io This was missing. I've been duct-taping n8n workflow logs with custom error handlers at each node boundary to get any visibility into where agents fail mid-run. Native OpenTelemetry tracing should make debugging multi-step agent failures significantly faster.

[Harel Asaf @RL_Asaf]

@charliejhills The serena MCP is underrated for this. Loading entire codebases wasted more context than it saved. Symbol-level reads cut my Claude token usage by ~40% on a recent project and actually improved output coherence.

[Harel Asaf @RL_Asaf]

@addyosmani AGENTS.md as a smell tracker is the right frame. Built a linter that scores repos on the broader signal set — not just whether the file exists, but the things agents actually rely on (entry points, test hooks, tool configs). https://t.co/im5vLbkn9n

[Harel Asaf @RL_Asaf]

@nikitabase Great breakdown. For the remote flavor specifically: threw together a Node.js starter with Bearer auth, rate limiting, URL versioning, and a Dockerfile + deploy guide so you skip the scaffolding and get to the actual tool logic. https://t.co/5efwcTeZBz

[Harel Asaf @RL_Asaf]

@n8n_io This is the missing piece for AI agent workflows in production. Every time I run Claude-based agents through n8n, the failure modes I can't catch are the ones that silently degrade, not crash. Trace-level visibility at the node execution level changes how you debug.

[Harel Asaf @RL_Asaf]

@ai_master007 The 47 agents number is the real flex here. Prompts matter but the gaps I keep hitting in n8n with Claude are state management between steps and tool call failures that cascade silently. What's your retry strategy when a sub-agent returns an empty response mid-workflow?

[Harel Asaf @RL_Asaf]

https://t.co/R9BMJ2YpkQ

[Harel Asaf @RL_Asaf]

@Siddharth87 Great breakdown of the harness internals. For anyone who wants a minimal version to poke at after reading this: a TS starter for the Agent SDK — Zod tools, async iterator loop, in-process MCP, clone + run in ~60s. https://t.co/R9BMJ2YpkQ

[Harel Asaf @RL_Asaf]

@curious_vii That first one's the worst — can't manage what you can't see. Built a standalone token counter for CC workspaces that shows what's eating the budget (CLAUDE.md, MCP, files, history), so you trim before you hit the limit, not after. https://t.co/sg0aKQBp5D

[Harel Asaf @RL_Asaf]

@JulianGoldieSEO This head-to-head format is the useful part — you learn more from one real bake-off than ten spec sheets. Built a little arena around exactly that: two agents run the same task and the winner's prompt gets revealed, so you see WHY one won: https://t.co/TjvcwxVcxR

[Harel Asaf @RL_Asaf]

@akshay_pachaar Sharp framing — it's a context problem, not capability. Hard part: you can't eyeball if YOUR repo carries the signals agents need. Built a linter that scores a repo's agent-readiness (AGENTS.md, docs, structure) — a number, not a guess: https://t.co/im5vLbkn9n

[Harel Asaf @RL_Asaf]

The memory layer is where I've seen most Claude-based agents collapse under real load. We had workflows passing sprint context correctly through 4 steps, then step 5 silently dropped the constraint bundle because token count crossed a threshold we hadn't monitored. The harness needs to know when its own inputs are getting truncated.

[Harel Asaf @RL_Asaf]

The manual hand-off pattern breaks hardest when task sequences grow longer than you tested. I've seen Claude workflows that pass 3-prompt chains in staging but silently accumulate context drift by step 7 in production. Dynamic workflows help, but state validation between steps still needs to be explicit.

[Harel Asaf @RL_Asaf]

The manual hand-off pattern breaks hardest when task sequences grow longer than you tested. I've seen Claude workflows that pass 3-prompt chains in staging but silently accumulate context drift by step 7 in production. Dynamic workflows help, but state validation between steps still needs to be explicit.

[Harel Asaf @RL_Asaf]

The prompt simplicity point is real but it hits a ceiling fast. In n8n workflows with Claude, the simple version works until you hit conditional logic across 4+ steps - then you need explicit state passing between nodes or Claude rewrites context it already resolved. Learned that after the 15th workflow rebuild.

[Harel Asaf @RL_Asaf]

Context7 is the one I keep coming back to. In production Claude workflows, stale docs fed through MCP cause silent hallucinations - the model answers confidently but from the wrong version. Serena helps on the code side but the doc problem is harder to catch without explicit validation steps between tool calls.

[Harel Asaf @RL_Asaf]

@aigleeson Good breakdown — "be the source it cites" is the real shift. The hard part is knowing where you stand today. Built a zero-dep Python CLI that scores any URL for LLM/generative-engine visibility, so you get the audit in one command: https://t.co/itoGY3e32l

[Harel Asaf @RL_Asaf]

@doncampbell That "it clicks" moment is the best part. Natural next step is code you can actually poke at — made a minimal TS starter for the Agent SDK: Zod tools, async iterator loop, in-process MCP, clone + run in ~60s. https://t.co/R9BMJ2YpkQ

[Harel Asaf @RL_Asaf]

@iannuttall The CLAUDE.md + MCP breakdown is the useful part — that's what quietly eats the budget. Built a standalone token counter for CC workspaces that shows what's hogging context (CLAUDE.md, files, history) so you trim before compaction hits. https://t.co/sg0aKQBp5D

[Harel Asaf @RL_Asaf]

@boringmarketer Nice workflow. To skip the manual scrape next time — built a zero-dep Python CLI that scores any URL for LLM/generative-engine visibility, so you get the audit in one command instead of prompting for it. https://t.co/itoGY3e32l