At a recent Foo Camp, several attendees expressed a common sentiment: they no longer want to consume someone else’s code as-is. Instead they would have an agent select the pieces they need, or extract patterns to reimplement (sometimes in a different language such as Rust). Jesse Vincent ran a full session on this topic; his Plaude AI notetaker summarized views that natural-language prompts, generation processes and developer skills may be overtaking executable source code as the primary shareable artifacts of software development.
What kinds of artifacts are becoming valuable?
Participants named several emerging artifacts: “context bundles” (the collection of specs, examples, prompts and supporting material given to a model), Agent Skills (a folder of instructions and references an agent loads for a task—Anthropic published this format and other agent products have adopted it), evals (test-case sets that define what good work looks like), and the agent trace (a log of what an agent tried, which tools it invoked and which context it used). The agent trace is described as the closest analogue to commit history for work that does not become formalized as code.
Speakers also warned of a tension: evals are useful only if their test cases remain unseen by the model’s training data. Nick Vincent pointed out the risk that large-scale data collection and sales to model labs could leak eval answers into training sets, invalidating what the eval measures.
Pressure on maintainers and community recognition
Jesse Vincent noted maintainers are often overwhelmed by uninvited, automated and low-quality contributions. Traditional mechanisms that gave contributors visibility and credit—pull requests, repositories, conferences—helped people connect and build reputations. As AI makes replication and generation easier, communities may need new ways to acknowledge participation and create career paths. Some interaction has moved to Discord and private messages, but structured and visible opportunities remain necessary.
Scientific computing was singled out as particularly vulnerable: one panelist warned that AI-generated forks and agent attacks could threaten carefully verified foundations such as NumPy and SciPy. Scientific workflows rely on battle-tested libraries and reproducibility; thousands of unreviewed alternatives could undermine reliability.
Historical parallels and the ‘‘unbundling’’ of knowledge
The author draws parallels with past technological shifts. Where once binaries were the traded objects, open source popularized sharing source code repositories and created new artifacts—commit histories, reputations and public repos that served as résumés. Today, that unbundling appears to be happening again: instead of shipping a bundled library that encodes many problems, a repo may increasingly serve as a source of context for solving a specific, on-demand problem.
A concrete data point mentioned: Stack Overflow’s new-question rate peaked at roughly 200,000 per month around 2014 and was already declining before ChatGPT; that decline accelerated sharply after November 2022. The answers and community knowledge haven’t disappeared—they’ve been absorbed into models, which now synthesize responses and divert traffic from the original knowledge creators. New mechanisms are needed to circulate value back to creators.
The article also recalls how other artifacts rose to prominence in earlier revolutions: domain names and web pages in the early web, the click and engagement metrics in advertising and social media. By analogy, the stream of what people ask their agents and the traces agents leave could become the signal around which AI-era coordination and markets form.
Experiments and emerging platforms
Multiple projects are experimenting with ways to share and coordinate the new artifacts:
- Duncan Davidson has been working on the idea of a home for personal context.
- The Open Memory Protocol initiative aims to standardize mechanisms for exchanging context.
- Sam Schillace is collecting agent-building patterns; one pattern, “Anchor on Canonical Names,” stresses referencing canonical names as compact pointers to learned methodologies (a theme also explored in the O’Reilly Expert Intelligence MCP).
- New efforts to reinvent GitHub for agentic workflows include Cursor Origin and Entire (the latter founded by former GitHub CEO Thomas Dohmke). Entire stores intent alongside code and uses Trails—an agent-driven review system that logs session traces and automated evaluations so maintainers can merge without reading every line of a diff. Origin describes itself as an agent-first Git hosting platform with structured semantic diffs and machine-readable merge states.
- Anthropic described an internal replacement for Git called Scion in a detailed public write-up.
- Common Fabric is building a system premised on disposable, agent-generated software patterns that carry user-associated data and rules for use; their vision expects agents to generate small reusable pieces that anyone’s agent can fetch or reproduce on demand.
These projects illustrate both possibilities and limits of a model in which context and specs are distributed and code is regenerated on the fly.
Trust, standards and the risk of abuse
The article invokes William Cronon’s history of Chicago’s grain trade to illustrate that new tradeable artifacts require trust-enabling mechanisms (standards, receipts, oversight). Without governance, receipts or artifacts can be counterfeit or misused. Likewise, AI-era context artifacts and traces invite both mistakes and bad actors.
Early missteps are normal: the author recalls the spreadsheet revolution, where spreadsheet models became widely reused but were only as good as their assumptions. Similarly, poorly constructed or misused context bundles and evals will produce failures until the field develops robust practices.
What should communities prioritize?
The core challenge for practitioners and open-source communities is to identify which new artifacts are worth preserving, sharing and governing: spec-driven programs, well-formed conjectures or questions, eval holdout sets, canonical names, agent traces, and context bundles. They must also design incentives and institutions for credit, attribution and compensation. Not all software will become agent-generated—large SaaS, distributed binaries and explicit source code will remain important—but there is room for new, lighter-weight shared artifacts.
Closing perspective
The author recommends a historical lens: new valuable artifacts tend to appear well before the institutions that coordinate them. GitHub and Google each took years to become their ecosystems’ coordinating institutions. The current moment is similar: context artifacts and agentic traces are emerging, while coordinating institutions and measurement systems lag behind. Those who recognize and design for the new artifacts—measurement, trust and exchange mechanisms—will shape the AI-era software economy.
The article is an invitation to conversation: how do we preserve and share starter contexts, protect eval holdouts, attribute and pay for agent activity, and build trustworthy institutions around these new knowledge objects? The questions affect software and many other fields where AI integrates into workflows.
