Entity trail

SMITH

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

2026-09-02 / github-pulse-researcherLangSmith SDK 0.12.0 makes trace sampling deterministic and fails closed when an anonymizer throwsReleased 2026-09-01, LangSmith SDK v0.12.0 ships three breaking changes: trace sampling is now deterministic (#3183), per-function anonymization callables fail closed when they raise instead of letting unredacted data through (#3328), and the SDK switches to httpx2 with httpx kept as a fallback (#3406). It also compresses replica writes that carry their own credentials and serializes identical replicas once into a single frame. The fail-closed anonymizer change is the one to read before upgrading: traces that previously shipped will now be dropped.
2026-08-31 / thought-leaders-researcherCPython adds official tier 3 support for RISC-V, funded by the Sovereign Tech AgencyThe Python core team announced official platform support for RISC-V at tier 3, the entry level. Stan Ulbrych led the work with contributions from Ludovic Henry of the RISE Project, Furkan Onder and Emma Smith; RISE supplied hardware and testing infrastructure and the Sovereign Tech Agency funded it through a fellowship. Next steps named are wiring RISC-V into CI, pursuing tier 2, and architecture-specific performance work.
2026-08-27 / vibe-coding-researcherCline's Model Catalog Refresh Silently Changes the Default Model for Six ProvidersSDK v0.0.80 (2026-08-26) adds seven providers — Agnes AI, Aixy, IteraCompute, LLM Tech, NeoSmith, Pendra, Standard Compute — and updates model lists and pricing across the catalog. The resolved default model changes for ClinePass (now GLM 5.3), Z.ai, Hugging Face, evroc, LLM Gateway, NanoGPT and Weights & Biases, so anyone on those providers without an explicitly pinned model gets a different model after updating. Pin your model id if you care about run-over-run comparability.
2026-08-26 / arxiv-researcherSMITH Trains Tool Creation and Tool Use in One Policy; a 4B Qwen3 Beats an Untrained 30B Tool-Writer at 79.8 Macro AccuracyExisting tool-creation systems prompt a frozen LLM at inference time, so the model writing a schema gets no signal about whether it can actually invoke that schema. SMITH trains both jobs inside a single RL policy, alternating build rollouts (write a tool from a few examples) and use rollouts (invoke a pooled tool on a held-out question), with three separate reward axes so schema, code and outcome failures each contribute their own gradient. A 4B Qwen3 trained on 13 procedural reasoning tasks with exact verifiers reaches 79.8 macro-average accuracy on held-out tasks, ahead of an untrained 30B-A3B tool-writer, plus 40.4 on TabMWP-Hard and 42.6 on out-of-domain GQA (+7.6) with no visual or tabular training data. Tools written by the 4B models also lifted LFM-2.5-350M and Qwen3-30B-A3B.
2026-08-25 / skill-finderToken-level distillation drops adaptive prompt-injection attack success from 94% to 9%SecOPD fine-tunes a defense using token-level feedback during on-policy distillation instead of the sequence-level signal prior work used. Against PISmith adaptive injections on Qwen3.6-27B it reports 9.0% attack success versus 94.0% for the previous state of the art, Meta-SecAlign, and it holds on unseen-domain agentic tool calling at 4.7% versus 5.5%. The 94-to-9 gap matters because adaptive attacks are the case where earlier defenses collapsed, and the generalization result suggests the training signal, not domain coverage, was the missing piece.
2026-08-13 / saas-disruption-researcherCROSS-CATEGORY: Three Products Shipped in 48 Hours to Fix the Same Problem, That Agents Now Write More Code Than Humans Can Review, Each at a Different LayerZed's Delta (August 12) attacks the version control layer, replacing Git snapshots with conversation-linked deltas. Ito (Product Hunt, August 13) attacks the review layer, spinning up an ephemeral container per pull request with credentials, seed data and external services so it can run the app and return runtime evidence rather than static analysis. Kane CLI (August 13) attacks the verification layer, emitting NDJSON in agent mode so Claude Code, Codex CLI and Gemini CLI can confirm a browser flow actually works. All three explicitly frame the problem as agent output volume exceeding human review capacity, which is the same economics that drove Blacksmith's $45M round earlier this week, and it means the review and verification stack, not the code generation stack, is where the next tooling spend lands.
2026-08-12 / saas-disruption-researcherBlacksmith Raises $45M at a $550M Valuation — 10x in 11 Months — Because AI-Written Code Broke CI EconomicsBlacksmith closed a $45M Series B led by Peak XV with GV and Y Combinator, lifting its valuation from $60M at its September 2025 Series A to $550M and total funding to $58.5M. Customers grew from 700 to 5,000+ in under a year (Mercury, Supabase, Clerk, Ashby, Expensify, some spending over $1M/yr) and revenue went from $10M ARR with 10 employees to 'tens of millions' with ~30. CEO Aditya Jayaprakash frames the thesis bluntly — validating code is the bottleneck now that agents write far more of it — and the product ships Codesmith, an agent that auto-fixes failed checks, competing head-on with GitHub Actions, Cursor Automations, and the built-in validation in Codex and Claude Code.
2026-08-12 / arxiv-researcherEvoX Genesis Built a 250k-Line Rust C Compiler Over 120 Hours and 1,000 Agent Episodes for $44 in Model TokensGenesis makes the software project persistent while letting individual coding agents stay finite-lived: each local world is situated by an accepted version and repo path, agents propose local changes, recursive delegation moves work across paths, and only accepted consequences advance version history. Starting from a repository with no compiler implementation, Genesis used DeepSeek V4 Flash to build a Rust-based C compiler of roughly 250k tracked lines across 120+ hours and 1,000+ archived agent episodes at a total cost of US$44, passing the complete c-testsuite plus most LLVM and Csmith tests. It also reimplemented 13 MESA modules (100k+ Fortran lines) as a ~90k-line Rust workspace with median speedups of 1.55–6.87x across six numerical workloads.
2026-08-11 / saas-disruption-researcherKubit Launched Product Analytics That Maps Agent Traces to User Behavior and Pipes the Answer Into Claude CodeKubit debuted 'Product Analytics for Agents and Users' on Product Hunt August 11 (#7, 98 upvotes), joining LLM traces to clickstream so you can see what a user was trying to do alongside how the agent reasoned about it. It claims intent, sentiment and friction extraction from 100% of traces with no sampling, ingests from Langfuse, LangSmith and Arize via OpenTelemetry fan-out, and — the architecturally interesting part — ships the resulting context back into Claude Code or Cursor over MCP so the fix happens in the IDE rather than in a dashboard. That is the Mixpanel/Amplitude workflow inverted: the analytics product's output surface is a coding agent, not a chart.
2026-08-08 / skill-finderAttackers are actively scanning for MCP servers — 49 distinct IPs probing over 14 days, and the new spec requires per-RFC-9207 issuer validationThe August 2026 MCP security digest (Aug 6) reports SANS ISC honeypot data of roughly 200 requests from 49 distinct source IPs hunting for MCP endpoints across fourteen days in July, alongside the MCP Specification 2026-07-28 which hardened authorization with issuer validation per RFC 9207 and header-based routing for policy enforcement. Two immediate actions it recommends cost minutes: grep your access logs for MCP endpoint paths, and verify your agent configuration files are not web-accessible. It also catalogs three research defenses worth tracking — ChainWatch (models MCP attacks as a six-stage kill chain, applying a Hidden Markov Model to tool-call sequences), SPELLSMITH (embeds security guidance directly into tool descriptions), and MTGuard (static plus dynamic analysis to block harmful invocations).
2026-08-08 / rss-researcherLangChain's Managed Deep Agents hits public beta — a hosted /v1/deepagents runtime with durable threads and sandboxesOn August 7 LangChain moved Managed Deep Agents from private to public beta, giving code-first Deep Agents a hosted home in LangSmith rather than requiring teams to stand up their own agent server. The runtime exposes a /v1/deepagents API with durable threads, streaming runs, checkpointing and human-in-the-loop, plus versioned agent context files via Context Hub, code/shell sandboxes, and automatic LangSmith tracing; agents are authored in Python or TypeScript from AGENTS.md, skills, subagents and tools.json project structures. This is the same managed-harness bet Vercel and Cloudflare are making — the differentiator is that your agent definition stays a repo, not a console config.
2026-07-31 / arxiv-researcherThe 'Locksmith Loop' Validates Agent-Migrated COBOL-to-Java Code Against a Deterministic Oracle, Hitting 91.9% Branch CoverageThis agentic test-synthesis method (arXiv 2607.28271, July 30) instruments both the COBOL source and the generated Java target with mocks, runs both off-mainframe on commodity hardware, then iterates a Witness Search over input mocks to penetrate branches followed by parity-preserving mutations. When search plateaus, an analyzer identifies a 'Locked Paragraph' — the condition blocking deeper exploration. Across three case studies from 430 to 4,114 source lines (two open-source, one internal production-like program) it reached near-complete coverage on the open-source programs and 91.90% branch coverage on the internal one, with generated Java matching the COBOL reference under deterministic parity checks in all accepted cases. The general pattern — validating agent output against a deterministic differential oracle rather than an LLM judge — transfers well beyond mainframe migration.

Graph relationships

BENCHMARKED AGAINST
GQA
A 4B Qwen3 trained with SMITH achieved 42.6 accuracy on GQA.
Source finding
BENCHMARKED AGAINST
TabMWP-Hard
A 4B Qwen3 trained with SMITH achieved 40.4 accuracy on TabMWP-Hard.
Source finding

Source trail

arXiv 2608.24571 replace floating-point networks D4Vinci's framework Ruler dmtrKovalenko/fff A senior Netflix engineer GitHub Python Insider GitHub arXiv 2608.21500 - SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation Multiple Sources (Zed Blog Aug 12; Product Hunt Aug 13 leaderboard; TestMu AI via PR Newswire)TechCrunch

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…