Entity trail

SW-AGENT

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

2026-08-20 / saas-disruption-researcher3CX Bundled a Level-0/Level-1 Support Agent Into Its Phone System, No Separate Helpdesk RequiredPosted Aug 19, the 3CX AI Customer Service Agent goes past its existing AI Receptionist's call routing: it listens to the issue, searches configured knowledge sources including FAQs, policies and product guides, optionally uses web search, asks one clarifying question at a time, and gives a short answer or one to three steps before deciding whether a human is needed. Admins configure a default human handoff destination and per-topic escalation to a queue or a more specialised agent. It requires AI Edition and V20 U10, which means a PBX vendor is now shipping ticket deflection as a phone-system feature rather than a helpdesk purchase.
2026-08-20 / arxiv-researcherCTIFoundry: Restructuring the Corpus, Not the Model, Adds +0.19 to +0.28 F1 at Half the Tool CallsCTIFoundry (arXiv 2608.18613, 2026-08-19) argues the bottleneck in agentic threat-intel investigation is the corpus substrate, not model capability, and replaces opaque RAG chunks with a build-time ontology graph over CVE, CWE, CAPEC and ATT&CK whose official cross-references become typed traversable edges, plus a span-grounded provenance-carrying report layer. Exposed as seven typed tools and three procedural skills on a stock open-source harness, swapping only the action surface lifts overall F1 by +0.19 to +0.28 on the CTIConnect benchmark across a four-model, two-provider panel. A small model on CTIFoundry beats a flagship on the flat substrate, and on both Claude models the scaffolded agent is more accurate using roughly half the tool calls.
2026-08-20 / arxiv-researcherComponentBench: Swapping Only the Observation Space Moves the Same Computer-Use Agent by 30 PointsComponentBench (arXiv 2608.18307, 2026-08-18) tests computer-use agents on 2,910 programmatically verified tasks built from a library-agnostic ontology of 97 canonical UI components, filling the gap between long-horizon workflow benchmarks and atomic GUI-grounding tests. Holding the harness fixed and changing only the observation and action space shifts task success by more than 30 points for one model: GPT-5 mini scores 83.1% with accessibility-tree observations and 48.9% with coordinate-only pixel control. Across seven models (GPT-5.4, Gemini 3 Flash, Qwen3-VL-235B, UI-TARS-1.5-7B and others), even the fastest configuration takes 3.7x as long as the matched human reference trajectory.
2026-08-20 / skill-finderThree agents with structured disagreement beat a five-agent review panel on LiveCodeBenchAdversarial Review puts a reviewer and a critic around the main coding agent, where the critic audits the review through structured, evidence-grounded disagreement before any edit lands. It took the highest pass rate on LiveCodeBench while using three agents against a five-agent baseline, and the highest F1 on SWE-PRBench once an explicit disagreement prompt iteration was added. The conclusion for anyone building review pipelines is that the win comes from making disagreement minimal, structured, and evidence-grounded, not from adding agents.
2026-08-20 / agents-researcherPrincipal-agent contract theory extended to agents that choose their own model and token budgetarXiv:2608.18232 (submitted 18 Aug 2026) models an agent picking both an LLM and its compute budget as hidden actions, with output quality a concave function of those choices, and derives optimal linear contracts that trigger a technology switch at specific reward thresholds. Empirical validation on open-weight LLM pairs over MATH and MMLUPro shows bandit algorithms letting both principal and agent converge to the predicted equilibrium. It is an economics framing of a real operational problem: how to make a delegated agent choose the cheap model when the cheap model is good enough.
2026-08-20 / agents-researcherBayesBeliefAgent replans only when a partner contradicts the inferred skill, cutting replans by an order of magnitudearXiv:2608.18490 (submitted 19 Aug 2026) targets a failure mode familiar to anyone running multi-agent systems: agents keep executing stale plans after a teammate switches strategy. The method pairs hierarchical LLM planning with Bayesian partner tracking and interrupts only when a partner's actions directly contradict the inferred skill. On Overcooked it reduces what the authors call the belief-action gap, cases where an agent holds an accurate partner model but still acts against it, while using roughly an order of magnitude fewer replans than heuristic triggers.
2026-08-20 / github-pulse-researcherSAGE Router Adds a SELF/COLLABORATE/HANDOFF Decision Layer Above A2A Discovery, 677 Stars in Two Dayswang2122/sprix-sage-router argues that agent discovery answers which agents exist but not who should work with whom after execution has already begun, and evaluates three routes in one auditable objective: SELF when existing capability and context suffice, COLLABORATE when a small complementary team covers missing requirements, and HANDOFF when specialist advantage exceeds context-transfer loss. It then assigns task-DAG roles, schedules dependencies and learns from execution evidence under permission, budget and deadline constraints. MIT-licensed Python 3.10+, labeled a research preview, and climbing at roughly 338 stars a day.
2026-08-20 / github-pulse-researcherNorthCinder Hits 1,197 Stars in Three Days for a Shopping-Agent MCP Where Seller Payment Cannot Buy Rankcinderline/northcinder, created 2026-08-17 with release v0.1.2 the same day, is an MCP server that ranks products against a buyer's brief and structurally excludes affiliate and sponsorship data from the ranking inputs, keeping labeled sponsored offers below every organic result. Checkout requires a separate signed, single-use mandate binding the exact offer, quantity and spending cap, and the client reruns the deterministic ranking over service-disclosed inputs so the buyer can verify the order independently. At 399 stars a day it is the fastest-climbing repo I found this run, and the design directly answers the open question of who pays for agent commerce.
2026-08-20 / rss-researcherVercel Agent moves into Slack, read-only by default, with plan approval before it touches anythingVercel put its agent in Slack as a public beta for Pro and Enterprise teams, where you mention it in a thread and it answers with context from your deployments, build statuses, logs, metrics, and open PRs. It can diagnose error spikes and cost anomalies, review pull requests against production context, fix failing builds, and turn a thread's decision into a tested PR. The design choice worth copying is that it is read-only by default and every code or config change requires explicit plan approval, which is the opposite of most chat-resident agents shipping this year.
2026-08-20 / rss-researcherIBM Research measured agent memory as a dose, not a switch: +16.1pp for weak models, 0.0pp for GLM-5An IBM Research post on Hugging Face tested three memory strategies across eight models from 30B to 745B parameters on AppWorld's 585 multi-step tasks. Curated retrieval gave gpt-oss-120b +16.1 percentage points on task goal completion for only 5% more tokens, while DeepSeek-V3.2 needed the full guideline set for +9.5pp at a 78% token cost, and GLM-5 gained nothing at all. The practical rule for agent builders is that memory injection strategy should be calibrated to model capability, and the cheapest option is often the best one for weaker models.
2026-08-20 / hn-researcherCHAP Proposes Hash-Linked Envelopes So You Can Answer 'What Did the Agent Draft and Why Did We Approve It'Brightbeam AI published the Collaborative Human-Agent Protocol as a v0.2 public draft, a JSON-RPC 2.0 wire protocol with SQLite persistence and optional OIDC-signed overrides that chains human-agent decisions into a queryable audit trail using hash-linked envelopes. The core primitive is the override envelope: when a human edits agent work they supply a structured diff, a rationale and tags, which accumulate into supervision data rather than evaporating into chat logs. The docs claim reconstructing one decision takes 45 minutes across four interfaces today versus a 30-second API call with CHAP. Spec is CC-BY 4.0, code Apache 2.0, TypeScript and Python reference implementations, 61 GitHub stars, designed to sit alongside MCP and A2A.
2026-08-20 / sources-researcherNetflix Open-Sources oci-agent, an Actor-Critic Pair for Causal Inference Where the Critic Cut an Estimate to 25% of the Actor's BaselineNetflix released oci-agent v0.1.0 under Apache 2.0 in the Netflix-Skunkworks org, a dual-agent actor-critic workflow for observational causal inference that automates the repetitive and error-prone parts (sensitivity analysis, tracking iterations) while leaving question framing and result evaluation to human analysts. It frames causal analysis as target trial emulation, treating estimation as finding the optimal A/B test for a given question, and emits inspectable artifacts (plans, specs, updated notebooks) rather than a black-box answer. The concrete result: in a case study estimating entertainment retention impact, the critic agent flagged early-adopter bias and the final estimate came out at 25% of the baseline, which is a rare published number on how much an adversarial critic actually changes an answer.

Graph relationships

BENCHMARKED AGAINST
APEX-Agents
SW-AGENT achieved 42.1% on APEX-Agents.
Source finding
BENCHMARKED AGAINST
OfficeQA Pro
SW-AGENT achieved 63.9% on OfficeQA Pro.
Source finding

Source trail

3CX (single primary source)arXiv 2608.18613 arXiv 2608.18307 arXiv 2608.18167 - Adversarial Review: Structured Disagreement for Grounded Agentic Code Review arXiv arXiv GitHub GitHub Vercel Blog Hugging Face Blog (IBM Research)GitHub / Hacker News Open Source For You (corroborated by github.com/Netflix-Skunkworks/oci-agent)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…