I deleted 90% of my harness
Rebuild the harness every six months and delete what models absorb—often down to one bash tool and inline scripts.
01 · The bitter lesson
The model eats the harness
Every model release invalidates a layer of prompt engineering and custom tool glue that engineers built to work around last year’s blind spots. Philipp Schmid invokes Rich Sutton’s Bitter Lesson: because post-training optimizes directly for task outcomes rather than obeying hand-wired tool trees, models learn cheaper paths and custom harnesses turn into dead weight.
“General methods that leverage computations are ultimately the most effective by a large margin”Philipp Schmid · I deleted 90% of my harness · 00:00:13
Rebuilding the Gemini agent harness from scratch every six months revealed that 90% of his custom file tools, planners, and memory helpers had become pure overhead—collapsing cleanly into one bash tool inside a sandbox where the agent writes inline Python scripts when needed.
“If the agent itself is a CLI, it can call itself using bash.”Philipp Schmid · I deleted 90% of my harness · 00:04:31
Across the summit, builders saw the same shrinkage from opposite ends. Justin McCarthy watches brownfield factory skills shrink without manual rewrites; Steve Yegge warns that refusal to prune turns yesterday’s scaffolding into “garbage mountain” (`00:01:59`); and Ishaan Dey urges teams to relax rigid constraints as intelligence scales (`00:02:57`).
“The skills are getting shorter. Harnesses are getting more general, of course, and a lot of this again is credit to the models themselves.”Justin McCarthy · Brownfield Factories · 00:08:09
02 · Tool surfaces
Swap API wrappers for semantic tools and primitives
Pre-classifying user intent and pre-fetching schemas used to feel like helpful optimization. At Sentry, Arsalan Bashir found those guardrails aged terribly across model upgrades—hiding information the model could actually use when investigating production issues.
“The first and obvious one was if this optimization kind of became a blind spot for the model.”Arsalan Bashir · Let the Agent Cook · 00:05:18
Consolidating overlapping per-API wrappers into clean, use-case-shaped tools—and pairing them with programmatic tooling where the model writes code directly against APIs (`00:03:01`)—produced immediate latency and cost gains.
“The agent only got four times faster, but also used 30 percent.”Arsalan Bashir · Let the Agent Cook · 00:08:07
Take that pruning all the way down to multi-agent coordination and even workflow graphs disappear. Lance Martin’s 1,000-agent math formalization ran with no bespoke orchestration DSL, Jarred Sumner replaced rigid workflows with long-lived subagents holding deep codebase context (`00:14:51`), and Simon Willison notes that Fable-class models brute-force hard problems once given clear tools and goals (`00:08:02`).
“Two different things: Git and a SQL database.”Lance Martin · Exploring Claude’s frontier · 00:02:06
03 · Runtime latency
Harden the runtime for 10× faster inference
When orchestration shrinks, execution speed becomes the real product constraint. Natalie Serrino cites June benchmarks showing why coding agents still break developer flow—forcing context switches far past the 10-second human attention boundary (`00:01:02`).
“From data from June, the median coding agent turn is 63 seconds.”Natalie Serrino · When AI Is 10x Faster · 00:02:57
Crossing 10× inference speed pulls agents onto the live critical path—such as blocking malicious uploads inline (`00:12:18`)—and immediately exposes a new bottleneck inside the runtime itself:
Counter-claim (bottleneck shift): Cutting inference latency 10× does not automatically make turns 10× faster. Tool execution balloons from 18% to 67% of wall-clock time, forcing builders to harden sandbox boot speed and tool pipelines next.
“And have 10x faster inference. We can see that 67% of that will be tool.”Natalie Serrino · When AI Is 10x Faster · 00:10:37
Teams are also cutting runtime overhead by routing routine classification away from heavy LLM loops—Cloudflare pairs System One models with a single LLM call for 94% fewer tokens, and Parallel’s grounded decision API answers in under a second at $1 per 1,000 requests (`00:03:01`).
“We shrunk our LLM calls from five down to one. It was about 94 percent less tokens.”Michelle Chen · cloudflare is a neolab · 00:14:06
04 · Organizational surface
One shared agent with public traces
Deleting brittle harnesses at the model layer only helps if teams stop reinventing private agent configs at the org layer. Zach Bruggeman shows that letting every developer tune separate dotfiles wastes days across a company (`00:03:29`). Ramp standardized on one shared Inspect agent (`00:05:14`) where every session is public by default—eliminating training decks because engineers watch how peers steer agents.
“Build tools that kill information silos.”Zach Bruggeman · Agent that builds the culture · 00:04:25
“It's actually easier to work in public than in private.”Zach Bruggeman · Agent that builds the culture · 00:05:38
Quinn Slack took that convergence to the entire company when spinning out Amp—replacing separate trackers and support queues with shared agent threads—while Utkarsh Shrivastava emphasizes at Netflix that the chat UI is merely the replaceable tip above durable company context (`00:02:11`).
“We got rid of all of the other tools. We got rid of all of the other things you got to check. Obviously, not Jira. No task tracking. We don't have a support system. Everything we do in the agent.”Quinn Slack · A Company Built Around the Agent · 00:08:04
05 · Counter-claims
Where deleting too much breaks
Deleting orchestration glue does not mean abandoning engineering structure. Three additional counter-claims—alongside Serrino’s tool-latency warning—draw the exact boundary between disposable scaffolding and durable systems:
1. Deterministic code beats prose rules: Lauren Tan warns that deleting tools only to stuff instructions into Markdown rules makes systems less reliable. Constrain agents in APIs, types, and lint rules first.
“Instead of adding another line to your rules or your agent's dot md, because it's far more reproducible, it's in code.”Lauren Tan · All Code Is Slop · 00:02:55
2. No single alien CLI wins everything: Thariq Shihipar rejects the fear that frontier models force humans to submit to one bare Bash paradigm; models form a broad “smooth interior” that works best when paired with linted, human-readable spec layers.
“Describe this kind of like fear that the models will just have this like way of doing things, and it will be weird. And alien to us, and we just need to accept it. You know, like we just have to like submit to this like foreign intelligence. And that's just like I don't think that's what happened, you know.”Thariq Shihipar · The Sweet Side of the Bitter Lesson · 00:01:40
3. Edge & vertical nuances stay: Michelle Chen notes Cloudflare’s initial bet that models would shrink to fit edge nodes missed trillion-parameter scaling, while Mihail Eric shows vertical AI still relies on shared chassis patterns filled with expert-guided domain workflows (`00:15:00`, `00:20:57`).
“Our hypothesis was that models were going to get smaller.”Michelle Chen · cloudflare is a neolab · 00:04:24
“there are principles and primitives that are true to any vertical company, vertical AI company.”Mihail Eric · Agentic GTM at Monaco · 00:08:33
06 · Playbook
Rebuild on a six-month cadence
The architectural shift across all 15 talks is an inversion of where engineering hours compound. Design choices tailored to yesterday’s weaker models actively throttle today’s models—so teams have to treat harness deletion as recurring hygiene rather than a one-time refactor.
“To create the most value, challenge your past assumptions, and rebuild your foundations to create good systems that scale with intelligence.”Ishaan Dey · Judgement is the Last Job · 00:04:35
For TheFocus.AI: Turn this into a repeatable client audit every model cycle: benchmark real tasks, delete custom routing and per-API wrappers in favor of
bash or coarse use-case tools, harden sandbox and tool-call latency, encode invariants in deterministic checks, and consolidate the team onto one shared agent with visible sessions.