Google DeepMind engineer Philipp Schmid says it works better after deleting 90% of the harness
Google DeepMind engineer Philipp Schmid shared "I deleted 90% of my harness (And It Got Better)" at the AI Engineer Code Summit 2026, streamlining the coding agent's tools from 10 to 20 specialized tools down to just one Bash, and reducing the system prompt from about 12,400 tokens to 574. He used Gemini 4 Argon to run four runtime frameworks on more than 350 of his own tasks; the score gap was within 6 points and tokens differed by 39%, but the cost per item ranged from $2.33 to $5.33, with Bash-only scoring the highest yet costing the most. He also compared Gemini 3.5 Flash with Gemini 4 Carbon, showing that the stronger the model, the better it can merge multiple steps into a single call.
Why read this
AI score 65
The original article provides a comparison of tokens and costs for different runtime frameworks on the same task, which can be used to judge the trade-offs of tool streamlining.
Topics
Body · AI translation
At the recently concluded AI Engineer Code Summit 2026, Philipp Schmid from the Google DeepMind team gave a talk titled "I deleted 90% of my harness (And It Got Better)", and a few days earlier he had also published a blog post, "Why Agents Are Soon Only Going to Have 1 Tool":
Philipp Schmid's blog posts are highly practical from an engineering standpoint, and I recommend following and reading them! AI Engineer Code Summit 2026 has also prepared a transcript for Philipp Schmid's talk:
https://ai.engineer/code/2026/talks/philipp-schmid
An overall interpretation of the two days of talks at AI Engineer Code Summit 2026:
Same result, but the path can differ by nearly tenfold in tokens
He used a post-training diagram to explain: the task, model, tools, compute environment, and grader are all present, but the model only receives a reward for the result, and no one specifies which tool order must be used to complete it. So the same repository change can either be broken into many specialized read and edit calls, or written as an inline Bash script and done in one go.
What he gave is an illustrative example: the task is to change amount in charge_customer to amount_cents, and update all callers. The specialized file-tool path took 10 rounds, with roughly 60,000 input tokens, and each round had to send the context again; the Bash path took 2 rounds, with about 7,000 input tokens, and 4 files were changed and the checks passed. He verbally called this gap about 10x; these tokens are estimates of per-round input.
Keep only one Bash, and put planning and memory into the file system
He said that roughly every six months he tears down and rebuilds the runtime framework for coding agents, asking how much is still needed. The 2025 generation had about 10 to 20 specialized tools, responsible for editing, planning, and to-dos respectively, and the system prompt also contained a large number of rules. In his most recent experiment, he kept only one tool: Bash.
Title: "A 1-tool agent: a computer instead of a toolbox". This Bash tool runs in a clean subshell, and its parameters are only the command, a description for the user, and async_after_ms (default 120,000, i.e. 120 seconds). After the command times out, it moves to the background and continues running, and the call returns the process ID and log path; the output keeps only the first and last 400 lines each, and the full text is saved separately to a file. Long tasks, reading logs, planning, and memory can all land on the file system. If the agent itself is a command-line program, it can also use Bash to invoke itself again, without needing to build another set of sub-agent tools. Reading, searching, editing, and testing are, in his view, all just commands.
But there is one practical limitation: printing an image in the terminal yields binary, and the model cannot see the picture. So in his evaluation, in addition to "Bash only", he also added a Bash plus view_media comparison group, where the latter sends images back in a form the model can see.
The system prompt is also getting shorter. In the PPT, each dot represents 100 tokens the model must read before the user prompt: in mid-2025, a coding agent with 16 tools was about 12,400; in 2026, his one-tool agent with Bash plus image viewing is 574. The numbers come from publicly available prompts and the system prompt plus tool definitions in local installations.
Scores are about the same, but cost and tokens are not
With the same model and the same task, different setups can lead to very different numbers of calls. He used Gemini 4 Carbon to run a problem in DeepSWE 1.1, and both sides passed: 8 structured tools took 47 model calls and 4.2M tokens; Bash only took 18 calls and 1.7M tokens. At the beginning, reads can also be parallelized, but once code editing begins, specialized editing tools change only one place per round.
He then ran his own set of more than 350 tasks covering coding, terminal, and office work, using Gemini 4 Argon across four runtime frameworks. Scores fell within 6 points, and tokens differed by 39%. The score is the average of three benchmarks, with the comparison baseline being Bash plus image viewing. These are his own evaluations, not a public leaderboard.
He explicitly said that each task was run only once, and that small differences in scores may just be noise, so this table cannot be used to rank accuracy. What really separates them is tokens, cache hits, and cost per task. Bash only had the highest score, reaching 74.9%, but it also had the most tokens and the lowest cache hit rate, at $5.33 per task, making it the most expensive instead. The most economical was Bash plus image viewing: 3.82M tokens, 89% cache hit rate, and $2.33 per task. The cost of structured tool A was about twice that. Fewer tools does not necessarily mean cheaper; without image-viewing capability, the tools saved may be eaten back by more tokens and lower cache hits. Looking only at scores, these four rows can almost be treated as the same tier.
The stronger the model, the more it can merge work into a single call
Deleting specialized tools does not automatically teach weaker models to merge steps. Using the Pi runtime framework, he connected Gemini 3.5 Flash, released about six months ago, and Gemini 4 Carbon to the default tools and to Bash only, respectively. For 3.5 Flash, the number of model calls was about the same with either interface: with Bash it also often used one command to read a file and then another command to modify it, with gains of only 0.95x to 1.7x, and some tasks barely changed. Carbon was different: with Bash only, tokens could drop to 1/1.4 to 1/3.0 of the original; each Bash call chained about 2.2 to 3.7 commands, while Flash had only 1.3 to 1.8.
There was a renaming task, and both sides followed the demo transcript. Carbon with default tools: 5 model calls, 15 tool calls, 17.2k tokens, basically reading each file once, then modifying it once, and finally running checks. Carbon with Bash only: in one model call, it used Python to modify 4 files, then ran checks, with only 3.5k tokens. Another slide summarized the difference as: a better model, switching from Argon to Carbon, with the same Bash-only framework, uses 34% fewer tokens per task; under the same model, the heaviest and lightest runtime frameworks can also differ in tokens by 39%. His judgment is that in 2027 people will more often ask: how many tokens does it take to solve a task. As models get stronger, old fine-grained tools and prompts may still be forcing them onto a slower path, so when a new model is released, it is worth tearing down the runtime framework again.
The next step may be the one tool that can run code
He plans to do the same experiment again in about six months. The current trend is a return to some kind of code mode: rather than writing free-form Bash commands, write TypeScript or Python directly. The PPT says: Bash handles files and processes, and everything else is left to exec(code). He also said he values MCP, especially now that it can return structured output, so the model can chain more actions in one round. Security is part of this design and cannot be solved just by having fewer tools: Bash calls are hard to inspect, but if there is a TypeScript abstraction layer in between, the actual operations executed can be checked; keys and policies stay on the host and do not enter the sandbox.
The closing three sentences are: give the model a computer, not a box of tools, that is, Bash and the file system; after a new model is released, tear down a layer and run your own evaluation; measure the tokens spent completing a task, not just accuracy. The slogan is: "Delete your code before the model deletes it."
AI Trends