Runkle ranks agent fixes, context first, then harness, then weights

Sydney Runkle posted eight takeaways from a conversation with Harrison Chase and Harry Stebbings on the Twenty Minute VC podcast, built on a single decomposition. An agent is a model plus context plus harness, and you can improve it by tweaking any of the three. Her ordering is that optimizing context is the highest ROI move, then the harness, then the weights, because context is where domain knowledge lives.

Sydney Runkle

@sydneyrunkle

X

memory is still massively unsolved, but also massively important

Sydney Runkle

@sydneyrunkle

X

the highest roi thing you can do is start by optimizing context (then the harness, then the weights)

That is a stated preference from experience, not a benchmark. Nobody in the post ran an ablation that moved the same task through all three layers and measured the difference. Treat it as a default starting point rather than a result.

The loop underneath it

The continual learning cycle she describes is run an agent, collect traces, aggregate learnings, fix something, measure the fix, repeat. The fix can land in any of the three layers, and the fixes look different, but the cycle does not change. The hard part, by her account, is distilling learnings from traces, which she calls incredibly domain specific.

She is blunt about what is missing there.

Cost comes last, visibility comes first

The cost section is the most concrete piece. The pipeline she gives is adding visibility into spend, putting controls in place, and only then pushing on agent optimizations. She says that is how LangChain reduced its own coding agent spend. No figure is attached to that reduction in the post, so if you want a number on what routing did to a coding agent's cost, Harrison Chase's earlier Open SWE post is where it came from, not this one.

Runkle also places model routing in the harness specifically, on the grounds that the harness is where the context needed for a routing decision lives. That is a design claim worth arguing with if you currently route at the gateway, outside the agent loop.

The rest of the list is directional. She says accuracy, cost, latency and security are the four things agent builders worry about, confirmed year after year in LangChain's dev surveys, though the post does not cite survey numbers. Open models, she says, are used for narrow tasks today, small parts of the agent loop or specific domains with post-training, and will power more core loops as they improve. And she expects both horizontal and vertical agents, with horizontal agents calling the vertical ones, which she says will force infrastructure changes because agents can do background work at a rate humans cannot.

For people building on it, the useful takeaway is the sequencing, not the conclusions. Spend visibility before optimization, and context before weights, are cheap orderings to adopt and cheap to abandon if your traces say otherwise.