Anthropic engineer: "Harness engineering is a really rare skill right now - and it's hard to know if you're even good at it." how he actually thinks about building the full system: • step 1 → the model. already smarter than we use it. "freeze development - you'd find... new capabilities for another 6-12 months" • step 2 → the harness. delete the steering. "we stopped using plan mode - the model thinks correctly" • step 3 → the agents. workflows let Claude build its own harness per task and spin sub-agents. what didn't fit in one context window now runs across many • step 4 → the context. "RAG is an anti-pattern. use grep." memory via files, clean context window, prompt caching. that's the stack most people optimize step 1. he optimizes all four. Watch the full video, then read the guide below on how to build an agentic system.show more

how does deleting steering impact real-world agent reliability?

Delete steering, and reliability drops fast - you’re not simplifying the agent, you’re removing the thing that keeps autonomy honest.

still amazed by how smart these engineers sound btw i'm actively studying harness engineering rn

Also, I’m glad we get to learn directly from the Anthropic team itself. Yeah, harness engineering is where I’m spending most of my time right now.

model was never the bottleneck

Maybe in the beginning of AI, but definitely not now.

Very clear explanation, thanks for sharing

the whole point is step 1 is done, everyone still tuning prompts is fighting last years bottleneck

yeah, sadly true

Thariq is always shipping alpha, great share bro !

Yeah, a real legend. I really loved the system he shared on the podcast.

Something new in my feed

yeah, podcast was posted less then 24-h ago

The harness insight is underrated but "delete the steering" only works when your codebase is small enough to fit in context. Past 50K lines the model needs explicit guidance or it drifts into fixing the wrong file. What repo size are you testing this approach on?

Learning directly from builders always gives way better insights than random takes

RAG as anti-pattern will get pushback, but it's technically sound. At 4k tokens, vector retrieval was necessary infrastructure. At 200k, you're adding latency and fragility for diminishing returns. grep scales.

The distinction between a stronger model and a better harness is the part most builders miss. Your long-form AI articles deserve one clear home where readers can browse the roadmap, newsletter and best systems—not just a Substack profile. Would a quick homepage concept be useful?

Deleting steering cuts down on prompt brittleness and over-constraint, letting the model use its native reasoning more freely. In real workflows we've seen it boost reliability on multi-step tasks (fewer hallucinations from forced plans), but it only works well with solid agent harnesses + clean context to catch drift. The engineer basically said they stopped babysitting and started trusting the model more - results improved.

skip the hype build the system

i’m trying grep before touching another vector database

The real insight isn't any single step. it's that harness quality is unmeasurable by most teams. You can A/B test a model. Almost nobody has a rigorous way to A/B test a harness, which is exactly why it stays a rare skill.

Optimizing all four steps instead of just the model is key

"RAG is an anti-pattern, use grep" is such a spicy line lol. Booking this one