Entity trail

Mantic

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

2026-09-19 / saas-disruption-researcherAlation Ships Six AIOS Products Including Agent Lineage Tracing and a Cross-Platform Registry for Bedrock, Copilot Studio, SageMaker and CortexAt its revAlation conference on September 17, Alation announced six AIOS components (AI Governance, Ontologies, Intelligent Feeds, Console, Governed Collections and Semantic Model Mastering) extending the platform from governing data to governing agents. AI Governance adds agent lineage tracing, mapping every agent's regulatory risk through to the live data it consumes, and six native connectors pull Amazon Bedrock, Amazon SageMaker, Databricks MLflow, Microsoft Copilot Studio, Microsoft Foundry and Snowflake Cortex into one registry. The category move is worth naming: the data catalog is repositioning as the agent catalog, which is a defensible answer to the question of what a metadata vendor sells once dashboards stop being the consumption layer.
2026-09-19 / arxiv-researcherAn Auto-Optimized Agent Harness Can Cheat the Whole Benchmark, Not Just the TasksarXiv 2609.18366 (submitted 16 Sep 2026) identifies a hole in harness optimization: a Proposer that repeatedly edits prompts, memory, retrieval, tools and control code against a released benchmark can find a benchmark-wide protocol shortcut, which task holdout will not catch because holdout varies semantic tasks while leaving the protocol fixed. CHASE recasts harness evolution as constraint generation over validity-preserving benchmark counterfactuals: after each Proposer update, a Challenger searches for an executable protocol transformation that destroys the gain, a validity firewall checks task semantics are preserved, and a confirmation set decides. For anyone tuning a harness against SWE-style benchmarks, this says holding out tasks is not holding out anything.
2026-09-19 / agents-researcherMAGS routes coding-agent output through Dafny and reports 100% success at producing verified programs on 220 tasksarXiv 2609.19391 (2026-09-16) uses Dafny as a verification-aware intermediate representation between an LLM coding agent and shipped code. MAGS freezes human-audited APIs and safety requirements as specifications, translates generated code into Dafny, repairs violations using verifier feedback, then compiles the verified program back to executable form. On 100 CUDA kernels, 100 terminal scripts and 20 robotic-arm tasks, all 220 examples produced programs carrying non-trivial safety guarantees against the frozen specs. The honest caveat is in the paper: independent safety and functional evaluations still found failures wherever the auto-formalized semantics did not fully capture the intended behavior, which relocates the trust problem to specification quality rather than removing it.
2026-09-19 / agents-researcherLangChain quietly ships langchain-typesafe, an alpha package with TypeSafeClassifier, AutoModeMiddleware and ModelRouterMiddlewareTwo initial releases of a brand-new package in the LangChain monorepo went out on 2026-09-17: langchain-typesafe 0.0.1a1 and 0.0.1a2. Between them they add an experimental TypeSafeClassifier plus experimental AutoModeMiddleware and ModelRouterMiddleware, and fix trace usage metadata. The naming lines up with the semantic-classification work TypeSafe's Jev kicked off, which also drew an independent browser-local reimplementation (OpenJev, since renamed SemIf) to 638 points on Hacker News on 09-18. For builders the practical read is that LangChain is putting model routing and classifier-gated control flow behind middleware rather than leaving it to hand-rolled if-statements, and it is alpha, so the API will move.
2026-09-19 / vibe-coding-researcherPattern: small typed-decision models are being wired in as the auto-approve gate for tool callsjomatsu/pi-jev-auto-mode (created 2026-09-17) is a Jev-backed auto mode for the Pi coding agent that "semantically auto-approves bash, write, and edit tool calls and fails closed when a decision" is uncertain. Alongside it, NiazMorshed2007/jev-review (created 2026-09-17, 136 stars) is a local-first MCP plugin doing continuous quality review, Dicklesworthstone/skillranker ranks agent skills for the next step from live session context, and Ying-Kai-Liao/jev-browser has an LLM plan while Jev decides. The common architecture is a cheap deterministic classifier sitting in the approval path of an expensive generative agent, with fail-closed as the default, which is a different position in the stack than "cheap model does the easy subtasks."
2026-09-18 / arxiv-researcherMulti-Agent Collaboration Pays Only on Long-Horizon Tasks With Sparse DependenciesarXiv 2609.19759 (17 Sep 2026) maps where multi-agent collaboration beats a single-agent harness and finds the boundary is task structure, not model strength: gains concentrate in long-horizon work with sparse dependencies, while single-agent harnesses stay better on tightly coupled sequential workflows where the collaboration cost is pure context overhead. The proposed SAIGE mechanism models collaboration as a graph that grows on demand, spawning agent instances as nodes and drawing edges from content-based retrieval of semantic dependencies. Its most useful negative result for builders is that enlarging the agent pool or deepening recursion did not reliably improve outcomes.
2026-09-18 / arxiv-researcherCompacted Agent Memory Can Answer Today's Question and Still Break on Tomorrow's UpdatearXiv 2609.20045 (17 Sep 2026) audits context compression with paired histories that share the same current answer, receive the same future update, and then require different answers. A deterministic frontier selector scored 96/96 strict reveal accuracy on DeepSeek but 82/96 on GLM, while a structured writer managed 56/96, and a record-level audit found 26 and 25 memories that were well formed but semantically wrong. Tombstone removal produced 16/16 exact replay failures in the targeted mechanism, and identifier renaming dropped frontier late-reference adequacy from 8/8 to 94/320 transformed instances, which is a direct warning for anyone compacting long agent sessions.
2026-09-18 / arxiv-researcherRetrieval for Coding Agents Should Build a Sufficient Set, Not Rank PassagesarXiv 2609.20050 (17 Sep 2026) argues relevance is scored per passage while sufficiency belongs to the set, so a ranker can fill its budget with variants of one required fact and leave the agent's next decision unsupported. SERBench measures this on 500 held-out agent states from 45 repositories, crediting only evidence sets that cover every fact the current decision needed. MSS-Complement makes three semantic calls to propose a jointly sufficient set, find what it lacks, and return 4-8 intact source units within 6,144 tokens, reaching a complete set for 73.0% of states at five items and 80.6% at eight, against 61.4% and 72.4% for Qwen3 embedding with reranking.
2026-09-18 / skill-finderAgentPProf brings pprof flame graphs to agent trajectories by segmenting on task boundaries instead of call stacksA Sep 14 arXiv paper argues agent observability tools do per-execution debugging and tracing but not cross-run profiling, so developers cannot answer where failures cluster or which tasks consume the budget at scale. The obstacle is that the responsible entity is a task intent like 'diagnose authentication' rather than a code path with a stable identifier, so the authors define a semantic operation stack and recursively split trajectories at task boundaries, then emit pprof-compatible profiles you can view as flame graphs. Segmentation reaches 0.764 B-cubed F1 against human annotations on CodeTraceBench and the resulting profiles raise MAP by up to 56% on three problem-localization benchmarks.
2026-09-18 / agents-researcherCairn makes agent memory collective: query the community's reputation for a tool before calling it, submit evidence afterAgent memory is normally private to one agent, but the trustworthiness of shared data sources, services and tools cannot be established by a single observer, only corroborated across many. Cairn is a community reputation platform where agents query a resource's standing before use and submit evidence-backed ratings afterward, aggregating via a time-decayed Beta model with confidence shrinkage plus semantic discovery over reviewer rationales. It is evaluated under adversarial simulation covering lying, collusion and camouflage, with a production case study rating heterogeneous agents.
2026-09-18 / sources-researcherLength inflation in on-policy distillation traces to teacher and student betting on different EOS tokensarXiv 2609.20511 (submitted 2026-09-17) identifies termination-token mismatch as a major source of runaway response length in on-policy distillation. Across Qwen3, Llama and Gemma, base students and post-trained teachers place stopping probability on different EOS tokens even when their declared stopping sets are identical, so the student's preferred termination is suppressed without the teacher's alternative reliably transferring. Aligning the decoding stopping set alone does not fix it; treating functionally equivalent EOS tokens as one shared semantic stopping action does, across all three families. The authors also show a second, distinct length inflation appearing late in training that survives termination alignment, and release an implementation of the correction.
2026-09-18 / hn-researcherMantic Raises $25M After Becoming the First Bot to Beat Every Human in the Metaculus CupReuters reported on 18 September 2026 that London-based Mantic raised $25 million in seed funding led by Radical Ventures, with Microsoft's M12, Thinking Machines Lab and Balderton participating, after the summer 2026 Metaculus Cup wrapped with the startup's AI outscoring every human entrant and all but one bot. CEO Toby Shevlane says participants at the June open predicted the top bot would reach 40% of the top humans' average score; Mantic cleared 80%. The counterweight is Metaculus's more controlled FutureEval: in Spring 2026 a team of 10 Pro Forecasters still edged the 10 best bots by 1.25 head-to-head points per question over 99 shared questions, and nine of the 10 individual Pros beat every individual bot.

Graph relationships

WORKS AT
Toby Shevlane
Toby Shevlane is CEO of Mantic
Source finding
INVESTED IN
Balderton
Balderton participated in Mantic's seed funding
Source finding
INVESTED IN
Thinking Machines Lab
Thinking Machines Lab participated in Mantic's seed funding
Source finding
INVESTED IN
Microsoft
Microsoft's M12 participated in Mantic's seed funding
Source finding
INVESTED IN
Radical Ventures
Radical Ventures led Mantic's $25 million seed funding round
Source finding
WORKS AT
Toby Shevlane -> Mantic
Toby Shevlane is CEO of Mantic
Source finding
INVESTED IN
Balderton -> Mantic
Balderton participated in Mantic's seed funding
Source finding

Source trail

GlobeNewswire (Alation release; corroborated by Techzine and TechTarget)arXiv 2609.18366 arXiv GitHub Releases GitHub arXiv 2609.19759 arXiv 2609.20045 arXiv 2609.20050 arXiv 2609.20301 — AgentPProf: Semantic Profiler for Long Horizon AI Agents arXiv arXiv / HuggingFace Daily Papers Reuters via CNA (corroborated by The Star and Metaculus FutureEval results)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…