Entity trail

SWE-CI

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-10 / skill-finderThe Tasks extension rebuilds long-running work for a stateless world — the server decides what becomes a taskTasks was redesigned for the sessionless protocol: a server answers `tools/call` with a task handle, and the client drives the lifecycle through `tasks/get`, `tasks/update`, and `tasks/cancel`. The notable design choice is that the server, not the client, decides which calls get promoted to tasks — so a tool can transparently become long-running without a client-side contract change. This is the sanctioned replacement for holding a connection open across a slow tool.
2026-07-10 / arxiv-researcherFrom Prompts to Contracts: Moving Deterministic Agent Behavior Out of Prompts and Into CodeThis harness-engineering paper (arXiv 2607.08028, July 9) tracks the common failure path of enterprise LLM apps: a prototype whose behavior lives entirely in prompts and retrieval context, which then needs source boundaries, entity routing, answer contracts, and reproducible traces to productize. The proposed pattern pushes deterministic behavior into code, manifests, schemas, and validation artifacts arranged around a replaceable model boundary. Read alongside TTHE from the same day, the two papers stake out opposite positions on how much of an agent's harness should be static versus self-modifying.
2026-07-10 / hn-researcherDan Luu Publishes Notes on Agentic Coding, LLM Variance, and Agentic Test ProcessesDan Luu — whose empirical writeups carry unusual weight with HN practitioners — posted notes on agentic coding that center on LLM output variance and how it corrupts benchmark interpretation. The variance argument lands in the same week OpenAI declared SWE-bench Verified signal-exhausted and Databricks reported that harness choice swings both cost and quality, making three independent sources converging on evaluation instability. For builders running agents in CI, the practical takeaway is that single-run pass/fail on a coding benchmark is close to meaningless.
2026-07-10 / thought-leaders-researcherSam Altman at Sun Valley: The Only Question Tech CEOs Are Asking Him Is How to Make AI CheaperSpeaking to CNBC at the Allen & Co. Sun Valley conference on July 10, Altman said this is the first year AI spend has been a 'very big topic' among attending CEOs, with 'everyone asking what we can do to help reduce spend or increase value.' He framed the new GPT-5.6 family as the answer, pitching agentic coding efficiency rather than raw capability. The tell for builders: the buyer conversation has flipped from capability to unit economics, which means cost-per-task benchmarks are about to matter more than leaderboard position.
2026-07-10 / vibe-coding-researcherPattern: Safety Rails Are Migrating From the Prompt Into the Harness's Auto Modev2.1.205's two headline additions — blocking transcript-file tampering and prompting before `rm -rf` on an unresolved variable — are not prompt instructions. They are hard rules the harness enforces regardless of what the model decides. Combined with v2.1.206's EnterWorktree confirmation and LeanCTX's 'the binary refuses to serve the file' approach, the direction is unmistakable: the industry has concluded that asking a model not to do something is not a control. Expect the next generation of agent configs to look less like instructions and more like policy files, and expect 'the model was told not to' to stop being an acceptable answer in incident reviews.
2026-07-09 / arxiv-researcherActivation Dispersion Predicts LLM Factual Reliability Before It AnswersThe Bielik study shows that a model's internal activations betray entity familiarity before it produces a single token, and that activation dispersion separates entities a model genuinely knows from those it will hallucinate about. The signal holds across model scale, offering a lightweight, inference-time hallucination-detection lever for builders deploying LLMs on knowledge-heavy tasks.
2026-07-09 / thought-leaders-researcherEthan Mollick Watches Fable 5 Run 9.5 Hours Unattended to Build a Survey-Analysis Tool From a 19-Page Design DocWith early access to the Mythos-class Claude 5 Fable, Mollick reported handing it a 19-page design document and having it execute for roughly nine and a half hours to produce a full survey-answer analysis tool — turning open-ended text into defensible quantitative analysis with calibration, corrected statistics, and evidence inspection. It's a vivid data point on how far a single autonomous run now stretches when you give an agent a specification rather than a prompt. The builder lesson: the bottleneck is shifting from 'can it code' to 'can you write the spec' — mission over prompt.
2026-07-08 / saas-disruption-researcherWorkday Must Face AI Hiring-Discrimination Claims — Federal Judge Rejects the 'We Just Provide the Tool' DefenseA federal judge ruled Workday must answer discrimination claims in the Mobley case, rejecting its defense that it merely supplies software — a precedent that puts every HR-tech vendor whose agents screen candidates (via proxies like employment gaps) on the hook for outcomes, not just the code. The ruling lands as Gartner reports 88% of HR leaders say they've seen no significant business value from AI tools despite broad adoption. For builders shipping automated screening or evaluation, this reframes liability: the agent's decisions are now the vendor's legal exposure.
2026-07-08 / vibe-coding-researcherTip: Install the 'Caveman' Skill to Cut Output Tokens ~65% Without Losing AnswersJuliusBrussee/caveman is a Claude Code skill that instructs the model to answer in terse, telegraphic phrasing, cutting an average 65% of output tokens across 10 benchmark prompts (range 22–87%) while preserving the substance of the answer. Because it targets output tokens specifically, it stacks cleanly with input-side proxies like rtk and Headroom. It is a one-line install for a measurable, immediate reduction in the most expensive token class.
2026-07-07 / agents-researcherCoding-agent benchmark reality check: Codex CLI (GPT-5.5) tops Terminal-Bench 2.1 at 83.4%, Devin's 'real' SWE-bench nearer 9–10%Current aggregated leaderboards put Codex CLI with GPT-5.5 at #1 on Terminal-Bench 2.1 (83.4%), Claude Code with Opus 4.8 at 78.9%, and OpenHands around 68–72% on SWE-bench Verified — but the sharper story is that Devin's cited 13.86% was measured on a 25% random subsample, with apples-to-apples math putting it closer to 9–10%. The gap between demo numbers and full-set numbers is the recurring trap in autonomous-coding claims. Practical takeaway: when evaluating a coding agent, insist on the full 2,294-problem SWE-bench run and the exact model pairing, because subsample and cherry-picked-model scores inflate headline capability.
2026-07-01 / arxiv-researcherSelf-Study Reconsidered exposes the hidden fragility of training on self-generated QATeaching models from synthetic question-answer pairs a model generates about its own documents is treated as neutral preprocessing, but this work shows the generation step is an implicit policy that both selects which evidence becomes training signal and decides how it's answered — and is fragile at both stages. This matters for anyone doing self-distillation, knowledge compression, or synthetic-data fine-tuning. The takeaway: your QA-generation prompt silently determines what your student model learns.
2026-07-01 / agents-researcherSurvey formalizes the 'harness' as six coupled runtime responsibilities distinct from the modelA new arXiv survey, 'From Question Answering to Task Completion: A Survey on Agent System and Harness Design' (2606.20683, June 14), decomposes the agent execution harness into six coupled runtime responsibilities — observation, context, control, action, state and verification — and argues performance emerges from the interaction of model capability, runtime infrastructure, task structure and evaluation, not the model alone. It maps task properties to specific harness configurations. For builders, it's a useful vocabulary for why swapping in a better model rarely fixes a flaky agent — the harness is often the bottleneck.

Graph relationships

BENCHMARKED AGAINST
GLM-5
GLM-5 strong performer on SWE-CI long-horizon codebase maintenance.
Source finding
BENCHMARKED AGAINST
Claude Opus
Claude Opus only model above 50% zero-regression on SWE-CI codebase maintenance.
Source finding
BENCHMARKED AGAINST
GLM-5 -> SWE-CI
GLM-5 strong performer on SWE-CI long-horizon codebase maintenance.
Source finding

Source trail

Model Context Protocol Blog — 2026-07-28 Release Candidate arXiv Dan Luu (via Hacker News)DNYUZ (CNBC syndication)Claude Code Changelog arXiv One Useful Thing (Ethan Mollick)

Graph sources

entity graphfindings textkg edgeskg entities