Entity trail

Agent Mode

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

2026-08-19 / skill-finderBenchmarked against workflows real users already pay for, the best agent finishes only ~30% end to endStartupBench inverts the usual benchmark construction: instead of researcher-invented tasks, the authors studied AI startup products with demonstrated market adoption, their workflows and their users, then translated those into complete deliverable-oriented tasks scored with fine-grained rubrics. Under a unified agent harness, even the strongest evaluated model completes only approximately 30% of the benchmark despite making substantial partial progress on many tasks. The named failure sources are the actionable part for anyone shipping agent products: complex instruction following and domain-specific expertise - meaning rubric-shaped task decomposition and injected domain context are where the remaining 70% lives, not raw model capability.
2026-08-19 / skill-finderAgents predict their own failure well (0.8847 AUROC) but cannot tell you which collaboration protocol will fix itAcross 4,181 competition math problems, this study compared direct solving, iterative self-correction, planner-executor-reviewer collaboration, and multi-agent deliberation, and found the routing question splits cleanly in two. Models reliably detect when they are about to fail - 0.8847 AUROC - but reliably fail at picking which specific protocol is worth its overhead. Practically: a simple confidence-based escalation reached 78% accuracy at 45K tokens versus a frozen router's 73.8% at 71.3K tokens, while a retrospective oracle hit 92.4%, leaving 18.5-28.9 points on the table in realistic settings - so gate escalation on self-reported confidence, but don't let the model choose the escalation shape.
2026-08-19 / skill-finderAttackers don't need your credentials - they can just spend your agent's resources: 84.06% success, best defense still leaves 55.11%Resource hijacking is a distinct attack class from data exfiltration: the attacker induces an agent to invoke, consume, transfer, or control high-value resources for the attacker's benefit without ever obtaining the resource or its credentials - think burning your compute budget, triggering paid API calls, or driving organizational workflows. ResourceHijackBench supplies 300 attack scenarios across six resource categories in isolated local environments that record actual resource consumption rather than simulated outcomes. Undefended, OpenClaw showed an average 84.06% attack success rate, and the strongest evaluated defense still left 55.11% - so budget caps and per-tool spend ceilings enforced outside the model, not refusal training, are the layer that matters here.
2026-08-19 / skill-finderMCP tools that write, not read, went from 27% to 65% of tool use - and measured defenses stop under 30% of attacksThis attack-surface survey of MCP, Skills, and tool calling reports that the share of deployed MCP tools which modify external state has climbed from 27% to 65% of tool use, which reframes the whole threat model: agents now act rather than read. Applied to blockchain execution, four properties - irreversibility, signing authority, continuous autonomy, and sequence-level composition - turn ordinarily recoverable agent-security failures into permanent loss, and the authors build a risk matrix tying each attack class to its amplifier, a representative mitigation, and the residual gap. The headline defensive number should worry anyone shipping write-capable MCP servers regardless of domain: measured protections stop fewer than 30% of attacks, and model-level safety refuses fewer than 3%.
2026-08-19 / skill-finderBind agent artifacts to content hashes and give the agent both parsed and native views: +8.3-12.1 points on OfficeQA ProStagedWorkspace attacks a failure mode most knowledge-work agent stacks have and don't measure: different components of the same agent referencing different versions of the same artifact. The fix is explicit version bindings between parsed records, review diffs, and native file content hashes, plus giving the agent dual access to both the parsed and the native representation. Dual parsed/native access improved Pass@1 by 8.3-12.1 percentage points over single-view alternatives on OfficeQA Pro and 4.7-9.2 points on APEX-Agents mean rubric scores, with SW-AGENT hitting 63.9% on OfficeQA (vs 29.3% published) and 42.1% on APEX (vs 25.5%) - the authors argue workspace state management should be treated as a first-class experimental variable, not plumbing.
2026-08-19 / skill-finderHow you lay out your repo changes prompt-injection success: highly modular workspaces measurably lower attack success rateThis is the first empirical study to treat 'workspace topology' - directory depth, codebase modularity, in-file injection position, and context framing - as an attack surface for coding agents that ingest third-party code with broad filesystem access. Across open-source repos spanning 10 languages and 6 engineering domains, testing three indirect-prompt-injection entry points against open-weight models on open-source harnesses, modularity changes significantly shifted attack success rate, with highly modular environments proving significantly harder to hijack; security cues placed in the workspace also moved ASR. Two takeaways for builders: repo structure is a security control you already own, and any coding-agent security benchmark run in a contaminated or non-representative workspace produces unreliable numbers.
2026-08-19 / skill-finderTrain the harness, not just the model: 6K examples of harnessed agentic RL lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%Agent Lightning v1.0 formalizes 'harnessed agentic RL' - the deploy-time harness (not the training engine) owns the environment interaction loop, and the trainer observes only sequences of LLM request/response pairs through an endpoint proxy. That inversion creates concrete engineering problems the paper enumerates and solves: retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, each of which materially affects training stability. In roughly 3,500 lines of code and with only 6K training examples on modest compute, RL took Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified (+14.6 points), and the complete workflow and training scripts are released - the disaggregated proxy architecture has already been adopted by verl Uni-Agent, AReaL 2.0, slime, and Polar.
2026-08-19 / skill-finderDeliberately mixing low-relevance same-domain items into context improved relevance accuracy by +0.077 - and six production patterns cut tokens 60-70%A practitioner framework from a production dashboard (meetings/email/chat -> structured work items) names six token patterns: context stratification, fetch-once/process-locally, schema-contracted prompts, token-aware fallback chains, semantic caching, and inter-agent communication compression; together they cut cold-load latency from a 3.5-10.5 minute baseline to 61-116 seconds with an estimated 60-70% token reduction. The counterintuitive result comes from a controlled study of 2,420 trials across 11 model configurations using 661 anonymized workplace items: holding the prompt at a fixed ten items, replacing some high-relevance items with same-domain low-relevance items improved the model's relevance-score concordance on the target items versus an all-high-relevance prompt. The 50:50 signal/noise condition beat the 100%-relevant condition by +0.077 (95% CI [+0.056, +0.098], Cohen's d = 0.49, Holm-adjusted p < .001, n = 220) - which means aggressive relevance filtering of retrieved context can actively hurt discrimination.
2026-08-19 / skill-finderAgent harness config is the most attackable phase, and detecting the risk does not stop it: >90% detection still coexists with high attack successHarnessRisk splits agent-harness safety into six operational phases (Harness Configuration, Capability Extension, Runtime Operation, State Persistence, Action Control, Incident Recovery) and runs 128 sandboxed cases that each pair a benign user objective with an adversarial instruction hidden in an untrusted workflow artifact. Across three harnesses, six models and 14 model/harness configurations, attack success spans 12.6% to 80.9% while utility stays at 75.0-97.6% - and Harness Configuration is the most vulnerable phase in all three harnesses, because attacks succeed by quietly altering security-sensitive parameters inside otherwise authorized workflows. The sharpest practical lesson: some configurations flagged the risk in over 90% of runs and still executed the attack, so 'the model noticed' is not a control - lock your settings/permission files as a separate trust boundary.
2026-08-18 / saas-disruption-researcherAtlas by WorkOS Landed on Product Hunt Selling an AI Coworker in Slack That Connects to 150+ Tools Out of the BoxAtlas hit #7 today (93 points), though WorkOS first announced it Aug 4. You @-mention Atlas in any channel, thread or DM, and it spans systems so a question crossing several tools returns one answer; teams define custom named agents by describing the job and scoping which tools each may access. The architecture is the interesting part for anyone building internally — Skills as reusable instructions, persistent Memory that survives conversations, and per-agent isolation with scoped integrations, server-side secrets and audit logs. That permission model is what an internal knowledge bot usually lacks, and it is WorkOS's actual home turf.
2026-08-18 / saas-disruption-researcherGrab Cut Mechanical Analyst Work From 44% to 30% With Agents — and Published the Five-Level Autonomy Model It UsedGrab's engineering team reported that the share of mechanical tickets handled by analysts (data prep, alerting, reporting) fell from 44% in February to 30% in June, with cycle times down roughly 33%. The share of requests answered with no human involvement rose from 53% to 67% for metric questions, 63% to 90% for data pulls, and 50% to 81% for SQL requests. The approach is worth copying more than the numbers: a five-level autonomy model defining how much of a workflow an agent may own, paired with certified data and explicit human oversight — the answer to 'how do you trust it' is scoped autonomy, not better prompts.
2026-08-18 / saas-disruption-researcherCROSS-CATEGORY: MCP Is Quietly Replacing the Vendor Dashboard — ElevenLabs, ZoomInfo and S&P Global All Shipped Connectors That Make Their Own UI OptionalElevenLabs launched a hosted MCP server in Claude (Product Hunt #4 today) that lets you create, inspect, update, duplicate and delete production voice agents — including revising a live system prompt and estimating LLM cost — without opening the ElevenLabs dashboard. ZoomInfo shipped a GTM MCP connector for Copilot Studio on Aug 11 exposing 500M+ contacts inside Excel, Word and Dynamics 365, and S&P Global expanded its Microsoft 365 Copilot integration on Aug 12 with an Excel connector for analyst modeling. Three unrelated categories — voice infrastructure, sales intelligence, financial data — all concluded within a week that the surface worth owning is the assistant's context window, not their own front end.

Graph relationships

RELEASED
Arena
Arena launched Agent Mode for users to run agents on real tasks.
Source finding
RELEASED
Arena -> Agent Mode
Arena launched Agent Mode for users to run agents on real tasks.
Source finding

Source trail

$37B annual run rate, up 123% year-over-year "Copilot code red"Gemini Enterprise Agent Platform $17B in bonds $1T market cap $10 trillion in AI infrastructure investment GitHub announced Nadella initiated a "Copilot Code Red"CodeBurn Nadella announced This open-source coding agent Researchers benchmarked

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…