Words
Writing
Essays, posts, and peer-reviewed papers on agents, retrieval, and planetary science, gathered from Medium, the Sourcegraph blog, journals, and here.
Essays
- Software factories are distributed systems
Reliability means preventing one layer's observation from becoming another layer's authority.
- I put Temporal around my software factory. Here's what it fixed, and what it didn't.
A durable work record is not the same thing as a durable procedure: what adding Temporal around a fleet of coding agents fixed, what it cost, and what it still has not proven.
- Scheduling an agent fleet: a practitioner's companion
A multi-agent orchestrator is an online scheduler, and its allocation layer, the part that prioritizes the work queue, is the piece most fleets leave on a default. The cheapest first build, the first change to make with the actual weights, context quality as a lever, and the rules that transfer from schedulers already running in production.
- The software factory as an observatory
Agent orchestrators are online schedulers, and astronomy has been fielding online schedulers for thirty-five years: what ZTF, Rubin, and the Deep Space Network teach about planning, proofs, fairness, and replaying your own nights.
- Operations research as a framework for software-factory scheduling
Technical report with its own literature review: operations research applied to an agent-fleet orchestrator. Dispatch, capacity, tier routing, acceptance gates, and architecture management mapped onto field-tested machinery from observatory scheduling, with a four-phase config-first adoption path.
- Running my own agent city on Slack
I drive a fleet of coding agents across a dozen projects from a Slack workspace: Gas City as the orchestration SDK, a Slack bridge I built in gascity-packs, and a Mayor that's really just a session aliased to a handle.
- Two retrieval systems write this site
The libraries, explorers, and digests on this site are produced by two retrieval systems exposed over MCP; the site is a static build in git, and git is the review gate.
- I used two multi-agent pipelines for everything I built this week. Here's what happened.
Running two multi-agent workflows — one for research and planning, one for execution and review — across a week of software projects, and what actually held up.
- Applying creativity research to agentic workflows
Borrowing divergent-thinking techniques from creativity research to push agent workflows toward structurally diverse solutions instead of the first obvious one.
- Why your coding agent keeps failing in ways you can't predict
Coding agents fail unpredictably because they are a complex system layered on another complex system, the codebase, and prompt engineering alone can't tame the emergent failures.
- Why "Agent Advocate" exists
On the Agent Advocate role at Sourcegraph, advocating for the infrastructure that makes coding agents work reliably in production rather than for better prompts alone.
- I couldn't find a good enough benchmark for large-scale software development, so I built one
Building CodeScaleBench — task design, verification, QA, and the infrastructure orchestration behind a benchmark for coding agents on large, multi-repo codebases.
- Rethinking coding agent benchmarks
Where today's coding-agent benchmarks fall short, and how to measure agent performance with stronger context retrieval at enterprise scale.
- How I use Amp (after 4 months and 6000 threads)
A working Amp practice after four months of heavy use — planning, context management, and verifying code across many projects.
- How I code with Amp on my phone
A secure iPhone-to-Mac setup with Blink Shell and Tailscale to run the Amp CLI and manage coding projects from a phone.
Engineering blog
- How we're using Sourcegraph and a Slack bot to detect vulnerabilities and react quickly
An automated Slack bot that triages security advisories, generates detection queries, and produces content artifacts to speed up vulnerability response.
- Why coding agents fail in large codebases (and what to do about it)
Five repeatable failure patterns for coding agents in large codebases, and why the fix is retrieval and code-search infrastructure rather than just better models.
- What it actually takes to run code intelligence in-house
The real engineering requirements and costs of building an internal code-intelligence platform equivalent to a commercial offering.
- Detecting supply chain attacks at scale with Deep Search
Using Deep Search and Code Search to find repositories exposed to poisoned LiteLLM packages, separating version-pinned code from genuinely at-risk dependencies.
- CodeScaleBench: Testing coding agents on large codebases and multi-repo software engineering tasks
A benchmark for evaluating how coding agents perform on enterprise-scale, multi-repo software development tasks across the full SDLC.
- Episode III: Revenge of the React vulnerabilities
New high-severity React Server Components vulnerabilities surfacing after incomplete initial patches — denial of service and source-code exposure — with remediation guidance.
- Fixing the React2Shell vulnerability in large and complex enterprise codebases (part 2)
Automating remediation of CVE-2025-55182 across many repositories with Batch Changes, the MCP server, and Deep Search — applying patches, tracking progress, and validating fixes at scale.
Guides & ebooks
- Running coding agents in enterprise codebases
An operational guide to deploying AI coding agents in large enterprise environments: context engineering, code retrieval, review pipelines, and human oversight.
- The Sourcegraph guide to surviving Big Code
Why universal code search across every repo, branch, and code host, on a unified platform, is essential to stay productive as codebases grow.
- CodeScaleBench: Benchmarking AI coding agents on real-world, large-scale codebases
A report on how AI coding agents perform against enterprise-scale codebases, finding that code understanding and context-retrieval quality are the primary drivers of success.
- Engineering Reliable Coding Agents
Evaluating and Operating the System Around the Model
Papers
All research & press →- JWST Spectroscopy of (142) Polana: Connection to NEAs (101955) Bennu and (162173) Ryugu
A. Arredondo, T. M. Becker, M. M. McAdam, A. S. Rivkin, S. Jarmak, I. Wong · The Planetary Science Journal
- Estimate of water and hydroxyl abundance on asteroid (16) Psyche from JWST data
S. G. Jarmak, T. M. Becker, C. E. Woodward, C. I. Honniball, A. S. Rivkin, et al. · The Planetary Science Journal
- Cassini UVIS solar occultation observations of Saturn's rings (data bundle)
T. M. Becker, S. G. Jarmak · NASA Planetary Data System (PDS)
- The science goals of Europa-UVS on NASA's Europa Clipper Mission
T. M. Becker, K. D. Retherford, G. R. Gladstone, et al. (incl. S. Jarmak) · 55th DPS Meeting (BAAS Vol. 55, No. 8)
- Experimenting with Large Language Models and vector embeddings in NASA SciX
S. Blanco-Cuaresma, I. Ciucă, A. Accomazzi, M. J. Kurtz, E. A. Henneken, et al. · arXiv preprint
- Apophis Specific Action Team Report
J. L. Dotson, M. Brozović, S. Chesley, S. Jarmak, N. Moskovitz, A. Rivkin, et al. · USGS Report
- Solar occultation observations of Saturn's rings with Cassini UVIS
S. G. Jarmak, T. M. Becker, J. E. Colwell, R. G. Jerousek, L. W. Esposito · Icarus
- The Adhesive Response of Regolith to Low-Energy Disturbances in Microgravity
S. Jarmak, J. Colwell, A. Dove, J. Brisset · Gravitational and Space Research
- Experimental and Numerical Investigations of Granular Dynamics in Microgravity
S. G. Jarmak · Ph.D. dissertation, University of Central Florida
- QUEST: A New Frontiers Uranus orbiter mission concept study
S. Jarmak, E. Leonard, A. Akins, E. Dahl, D. R. Cremons, S. Cofield, A. Curtis, et al. · Acta Astronautica
- CubeSat Particle Aggregation Collision Experiment (Q-PACE): Design of a 3U CubeSat mission to investigate planetesimal formation
S. Jarmak, J. Brisset, J. Colwell, A. Dove, D. Maukonen, S. A. Rawashdeh, et al. · Acta Astronautica