A common middle ground
Three open-source coding-agent systems that began with different design philosophies arrived at a similar set of building blocks, but in the versions examined none gave an outsider a tamper-evident way to verify its record, according to a preprint. The paper compares deepagents, pi and dsh - software frameworks around AI coding agents - and identifies five recurring elements: a commoditised loop, an append-only replayable session record, model quirks kept as data, progressive context disclosure and explicit extension seams. In ordinary terms, the systems share a basic way to run an agent, preserve its history, account for model-specific behavior, reveal context gradually and add capabilities at defined connection points.
The convergence did not follow a single route. The source histories pointed in different directions: deepagents subtracted authored scaffolding, pi accreted durable infrastructure, and dsh entered near the centre while literally reusing pi at one seam. The authors break the resemblance into three mechanisms - parallel discovery, diffusion and literal reuse. Because one of the cases borrowed directly from another, the shared shape cannot automatically be treated as evidence of three independent inventions.
History becomes the dividing line
The clearest contrast appears in the session record, the account of what the agent has done. Its durability forms a weak-to-strong ladder: deepagents has a weak view-based form, pi a medium append-only path, and dsh the strongest event-sourced form with a runtime assertion. An event-sourced form, in plain language, records a sequence of events rather than only a current view. The paper also notes that deepagents recovery still rewrites history and pi retains an older destructive compaction path, qualifications that keep the ladder from being a claim of perfection.
The paper also describes four seam specimens. Three ended in unrecoverable loss; the one recoverable specimen, identified as S1, crashed. The fault-line taxonomy is presented as small, analyst-directed and illustrative rather than as an independent test. These examples therefore show possible consequences at selected seams, not how frequently such failures occur.
The missing trust layer
That same question of trust sits outside the runtime. None of the three harnesses supplied a tamper-evident record that an outside party could verify without trusting the runtime. The authors identify external verifiability as the remaining boundary: the systems share several internal architectural elements, yet they do not provide a common, independently checkable account for outsiders. The study leaves that as an open design dimension rather than a solved problem.
What the evidence can support
The evidence combined pinned source code with commit, pull-request and issue archaeology, hands-on reproductions, and upstream confirmation of specific claims. The work was an explanatory, theory-building multiple-case study using literal replication logic: deepagents and pi grounded the candidate model, while dsh served as a later held-out check limited to independently instantiated elements. Parallel AI agents located and excerpted code, but every claim was then checked by hand. Reading depth was asymmetric, with dsh receiving one thorough pass plus a plugin experiment.
The result's reach is limited. The sample was N = 3 coding-agent harnesses, read at one point in time. The cases were not fully independent because dsh depends on pi, selection favored well-known source-available projects, and the five-element construct was induced from two grounding cases and checked against one more. A single analyst used AI assistance; dsh was read less deeply, and second-coder and inter-run agreement were deferred.
The document identifies itself as arXiv:2608.23953v1, dated 25 August 2026. Its message is therefore narrower than a field-wide verdict: the selected cases share a recurring middle form and a gap in external verification, while generalization to other harnesses and future revisions remains unresolved.
Paper data and sources
Original title: The Empire, Long Divided, Must Unite: Architectural Convergence in Three LLM Agent Harnesses
Authors: Dai Jiahong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text