· Mar 23, 2026
Siddharth Mishra-Sharma argues at Anthropic for long-running Claude in scientific computing
Markdown · JSON · Original source
Published March 23, 2026, Siddharth Mishra-Sharma describes using Code for a scientific computing project outside his core domain: implementing a differentiable cosmological Boltzmann solver. He argues that autonomous, long-running agent workflows can make progress on well-scoped tasks with clear success criteria, provided they have a plan, persistent memory, and a test oracle—a way to check progress against a reference or objective. For the solver, Claude reached sub-percent agreement with CLASS across various outputs after working from scratch over a few days. Mishra-Sharma also reports serious limits: the result was not production-grade, did not meet acceptable accuracy in every regime, and its test coverage had gaps. He presents the project as evidence that agent-driven development can compress work that might take months or years into days, not as a finished scientific tool. Watching the code’s incremental development also helped him learn the science, which he compares to reading lab notes from a “fast, hyper-literal postdoc.” His closing argument is that idle compute now has an opportunity cost: when projects have clear success criteria, not running agents may mean potential progress left undone.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Mar 23, 2026 (2026-03-23T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-07T19:03:23.251Z. First observed: 2026-10-23T01:41:11.804Z. Retrieved: 2026-10-29T16:00:10.684Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 29, 2026; originally published March 23, 2026. Mishra-Sharma reports that test coverage had gaps, including a period when tests covered only a single fiducial parameter point. The post describes one project; it does not establish how well the workflow generalizes to other scientific tasks.
Open question
How well would this workflow perform on scientific computing tasks whose success criteria are less clear or whose results lack a reference implementation?
Source-stated
Mishra-Sharma says long-running agent workflows can suit scientific computing tasks that are well-scoped, have clear success criteria, and allow occasional rather than continuous human oversight.
These are tasks where the work is well-scoped, the success criteria are clear, and human oversight can be occasional rather than continuous.
Source-stated
Mishra-Sharma describes using Claude Code to implement a differentiable cosmological Boltzmann solver outside his core scientific domain.
As a concrete example, I will walk through using Claude Opus 4.6 to implement a differentiable version of a cosmological Boltzmann solver.
Source-stated
In the example, Mishra-Sharma says Claude reached sub-percent agreement with the reference CLASS implementation across various outputs after working from scratch over a few days.
Claude worked on the project from scratch over a few days, reaching sub-percent agreement with the reference CLASS implementation across its various outputs.
Source-stated
Mishra-Sharma says the resulting solver was not production-grade and did not match CLASS to acceptable accuracy in every regime.
While the resulting solver is not production-grade (e.g., it doesn’t match the reference CLASS implementation to an acceptable accuracy in every regime)
Source-stated
Mishra-Sharma argues that agent-driven development can compress work that might otherwise take months or years into days.
it demonstrates that agent-driven development can compress months or even years of researcher work into days.
Source-stated
Mishra-Sharma says that watching Claude’s incremental progress helped him learn about Boltzmann solvers and their physics.
following Claude’s incremental progress and looking up what I didn't recognize turned out to be an effective way to osmose the science.
Source-stated
Mishra-Sharma argues that not running agents on projects with well-defined success criteria can mean potential progress is left on the table.
If you have the compute and projects with well-defined success criteria, every night you don't have agents working for you is potential progress left on the table.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 99bab15e-3444-9c05-5dd2-ccf3a8b78460.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T18:58:43.249Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: de5c4739811f3a31384fe6bf572b6630589037dd4d9147f9a705b38b9b360033. Source text hash: 2f24261308a3af53b3ed3b7bc050d5e67b937ea4d3ffaa4c0007b4aee7c583a4.
Revision and source-check history
Version 1 · 2026-11-07T19:03:23.251Z
Siddharth Mishra-Sharma argues at Anthropic for long-running Claude in scientific computing
Published March 23, 2026, Siddharth Mishra-Sharma describes using Code for a scientific computing project outside his core domain: implementing a differentiable cosmological Boltzmann solver. He argues that autonomous, long-running agent workflows can make progress on well-scoped tasks with clear success criteria, provided they have a plan, persistent memory, and a test oracle—a way to check progress against a reference or objective. For the solver, Claude reached sub-percent agreement with CLASS across various outputs after working from scratch over a few days. Mishra-Sharma also reports serious limits: the result was not production-grade, did not meet acceptable accuracy in every regime, and its test coverage had gaps. He presents the project as evidence that agent-driven development can compress work that might take months or years into days, not as a finished scientific tool. Watching the code’s incremental development also helped him learn the science, which he compares to reading lab notes from a “fast, hyper-literal postdoc.” His closing argument is that idle compute now has an opportunity cost: when projects have clear success criteria, not running agents may mean potential progress left undone.
Original source
· Article
Source details
Original source date: Mar 23, 2026 (2026-03-23T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-07T19:03:23.251Z. First observed: 2026-10-23T01:41:11.804Z. Retrieved: 2026-10-29T16:00:10.684Z.
Limits
Text as retrieved on October 29, 2026; originally published March 23, 2026. Mishra-Sharma reports that test coverage had gaps, including a period when tests covered only a single fiducial parameter point. The post describes one project; it does not establish how well the workflow generalizes to other scientific tasks.
Open question
How well would this workflow perform on scientific computing tasks whose success criteria are less clear or whose results lack a reference implementation?
Source-stated
Mishra-Sharma says long-running agent workflows can suit scientific computing tasks that are well-scoped, have clear success criteria, and allow occasional rather than continuous human oversight.
These are tasks where the work is well-scoped, the success criteria are clear, and human oversight can be occasional rather than continuous.
Source-stated
Mishra-Sharma describes using Claude Code to implement a differentiable cosmological Boltzmann solver outside his core scientific domain.
As a concrete example, I will walk through using Claude Opus 4.6 to implement a differentiable version of a cosmological Boltzmann solver.
Source-stated
In the example, Mishra-Sharma says Claude reached sub-percent agreement with the reference CLASS implementation across various outputs after working from scratch over a few days.
Claude worked on the project from scratch over a few days, reaching sub-percent agreement with the reference CLASS implementation across its various outputs.
Source-stated
Mishra-Sharma says the resulting solver was not production-grade and did not match CLASS to acceptable accuracy in every regime.
While the resulting solver is not production-grade (e.g., it doesn’t match the reference CLASS implementation to an acceptable accuracy in every regime)
Source-stated
Mishra-Sharma argues that agent-driven development can compress work that might otherwise take months or years into days.
it demonstrates that agent-driven development can compress months or even years of researcher work into days.
Source-stated
Mishra-Sharma says that watching Claude’s incremental progress helped him learn about Boltzmann solvers and their physics.
following Claude’s incremental progress and looking up what I didn't recognize turned out to be an effective way to osmose the science.
Source-stated
Mishra-Sharma argues that not running agents on projects with well-defined success criteria can mean potential progress is left on the table.
If you have the compute and projects with well-defined success criteria, every night you don't have agents working for you is potential progress left on the table.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 99bab15e-3444-9c05-5dd2-ccf3a8b78460.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T18:58:43.249Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: de5c4739811f3a31384fe6bf572b6630589037dd4d9147f9a705b38b9b360033. Source text hash: 2f24261308a3af53b3ed3b7bc050d5e67b937ea4d3ffaa4c0007b4aee7c583a4.
uncertain · 2026-11-07T18:58:43.249Z · anthropic/claude-opus-5.5
Publication: 99bab15e-3444-9c05-5dd2-ccf3a8b78460
attribution — pass: The reading correctly attributes the post to Siddharth Mishra-Sharma, an Anthropic Discovery team researcher, on Anthropic's site. The post's arguments and results are kept as his attributed claims. The evidence is a single authored post, not a listing.
In this post, Siddharth Mishra-Sharma, a researcher on the Discovery team, explains how to apply multi-day agentic coding workflows
summary — uncertain: The summary is faithful and keeps the author's force, caveats, and the 'fast, hyper-literal postdoc' voice. However, central terms are not explained for a general reader, and the definitions list is empty. The reading never says what a 'cosmological Boltzmann solver' is: code predicting the statistical properties of the Cosmic Microwave Background. It also does not explain 'differentiable' or that CLASS is the established reference Boltzmann code. Without this, the core development is hard to follow. Smallest repair: add a brief plain gloss of each term, or add tooltip definitions.
This is numerical code that predicts the statistical properties of the afterglow of the Big Bang—the Cosmic Microwave Background, or CMB.
date — pass: The evidence's originalPublished instant is 2026-03-23T00:00Z, and the reading's 'March 23, 2026' matches that UTC day at the source's precision. There is no hindsight framing, and forecasts stay attributed. The source is a publisher page, so no archive disclosure is needed. The page was edited after publication (dateModified 2026-10-09). The reading relies only on body text that is plainly integral to the original article.
support — pass: All claims match verbatim excerpts: sub-percent agreement within a few days, not production-grade, test coverage gaps, compression of months or years into days, learning from commits, and the opportunity cost of idle nights. 'Serious limits' slightly intensifies the source's 'somewhat clunky,' but the limits listed are all source-stated. The post-publication edit does not affect these passages, which are integral to the article body. The question follows from the source's own scoping to clear success criteria and is not a leading claim.
The agent’s development trajectory was somewhat clunky. For example, there were clear gaps in its test coverage
media — pass: No media was selected, and no verified media candidates were supplied, so no media is the correct choice.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore Anthropic