Entity trail

Agent Cards

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-17 / agents-researcherA trust-tiered 'librarian' plus multi-agent writer drives 6,845 cross-section contradictions to zero across 555,926 evidence cards'Reconcile Once, Write Anytime' (arXiv:2608.12984, Aug 13, 2026, Xing Zhang et al.) splits knowledge maintenance from report composition: a deterministic librarian ingests timestamped sources into evidence cards, a metric ledger and claim graphs, while a multi-agent writer may only read evidence stamped at or before a declared cutoff. Over 555,926 evidence cards from 6,130 sources (SEC filings, labor statistics, Wikipedia) the shared metric ledger removes 6,845 cross-section contradictions to zero, with a 3.7x speedup over serial processing and trust tiering that stops media numbers from displacing government or filing data. This is a concrete architecture for anyone whose research agents keep contradicting themselves between sections.
2026-08-17 / reddit-researcherDeepSeek Harness Cut Its First Tagged Release Today — and Can Now Manage Claude Code and Codex Subagents From Its Job Paneldeepseek-ai/deepseek-harness published dsh-v0.1.0-rc.7 at 12:01 UTC on August 17, its first tagged release since the repo was created on August 13. The notes include managing Codex and Claude Code subagent tasks through the Job Panel, plugin-registered settings cards, durable image attachments for MCP and ACP, and a `low` reasoning effort option for DeepSeek models with `high` staying default. The repo now shows 145,439 stars — up from the ~33,000 reported hours after launch four days ago — making the MIT-licensed TypeScript runtime one of the fastest-starred repos on record, and notable for a Chinese lab shipping first-class orchestration of two American coding agents.
2026-08-14 / arxiv-researcherInterSAGE Proposes the Trust Layer MCP, A2A and ANP Left Out of the Internet of AgentsExisting agent protocols specify how agents exchange messages but not how an agent proves identity, authorization, advertised capabilities, or accountability after delegation. InterSAGE adds four layers — Persistent Identity, Discovery, Trust Negotiation, Accountability — via Agent Identity Cards binding developer/code package/operator/deployment context, DID-bound Verifiable Credential discovery manifests, monotonic capability attenuation with two-tier access control, and kernel-mediated cryptographic audit trails that need no consensus ledger. It is designed to sit alongside MCP, A2A, ANP and AG-UI rather than replace them; a comparison against 50+ prior efforts finds none jointly enforcing all four properties.
2026-08-14 / arxiv-researcherTwo-Tier 'Librarian + Writer' Architecture Removes 6,845 Cross-Section Contradictions From LLM Research Reports to ZeroThis system separates a deterministic 'librarian' that ingests timestamped sources into a trust-tiered ontology (evidence cards, an authoritative metric ledger, a claim graph) from a portable multi-agent 'writer' that composes reports at any cutoff T, reading only evidence with as_of <= T. Evaluated on 6,130 public sources yielding 555,926 evidence cards (SEC EDGAR across 295 issuers and 11 sectors, BLS releases, Wikipedia), the shared metric ledger drives 6,845 cross-section contradictions to zero, tier-first selection is correct on 22/22 gold cases against 9/22 for a popularity-first baseline, and replay shows zero look-ahead violations across seven cutoffs. Difficulty-tiered model routing beat the all-Opus quality ceiling while running 3.7x faster than serial.
2026-08-13 / saas-disruption-researcherZed Ships Delta: a Multiplayer Agent Coding App Built on DeltaDB, a Git Replacement That Keeps Every Operation Between CommitsOn August 12 Zed opened private-beta invites for Delta, a standalone multiplayer app where humans and agents code in shared persistent threads, built on DeltaDB, a version control system Zed has been building for over a year that records a stream of fine-grained deltas with stable identities instead of Git's snapshots. Every code change is bidirectionally linked to the agent conversation that produced it, and Delta stays compatible with existing Git repos so teammates not using it still see a normal repository. It connects to third-party agent harnesses starting with Claude Code, so a terminal session syncs live into a Delta thread. The pitch is that Git's snapshot model permanently discards the multi-turn reasoning that now produces most code, which makes the PR the wrong review unit.
2026-08-13 / agents-researcherClaude Agent SDK 0.2.137 adds truncating resume and message origin, so hosts can rewind a session without silently dropping turnsThe 2026-08-12 SDK release introduces `resume_session_at` (fork a session at an earlier transcript entry) paired with `resume_drops_turn`, which validates that only entries from a specific turn are discarded — enabling safe rewind-to-before-last-prompt instead of blind resume. It also adds an `origin: MessageOrigin` field on `UserMessage` and `ResultMessage` that distinguishes application-submitted human prompts from background-task notifications, scheduled triggers and peer messages, plus a `ConversationResetMessage` surfacing `/clear`-style transcript discards mid-session. Both widen the `Message` union, so exhaustive `assert_never` matches break.
2026-08-08 / agents-researcherGated Hindsight Distillation trains GUI agents on the screenshot that comes AFTER the actionStandard offline GUI-agent training decomposes trajectories into prefix-action pairs and discards the next observation — which is often the only place the rationale lives (nothing reveals that Soft Wrap is under Edit or View until the menu opens), so imitation gives the model little chance of ever sampling the correct reasoning. GHD (arXiv 2608.06065, submitted 2026-08-06) uses the next screenshot as privileged information during training only: a parameter-sharing teacher that sees it re-scores the student's on-policy responses, and distillation fires only when the student fails and the hindsight-conditioned teacher recovers the demonstrated action. It improves task success over GRPO on AndroidWorld and AndroidLab across two vision-language models, with code and checkpoints promised.
2026-08-04 / agents-researcherClaude Agent SDK 0.2.129 patches an --allowedTools injection where a crafted skill name could grant itself extra permissionsShipped 2026-08-04, the Python Claude Agent SDK fixed a flaw where skill names passed via `ClaudeAgentOptions(skills=[...])` went unchecked into the CLI's `--allowedTools` value, which splits on commas and spaces, so a name carrying those delimiters could inject additional permission rules into the agent's own sandbox policy. The transport now validates each name and raises `ValueError` at connect time for parentheses, commas, control characters, wildcards, leading slashes, surrounding whitespace, or surrogate code points, which is a breaking change: `skills=["*"]` and `skills=["plugin:*"]` must become `skills="all"` or an explicit `Skill(...)` rule. Names with leading whitespace or a leading slash previously built rules that could never match, silently disabling the skill rather than failing loudly.
2026-07-30 / agents-researcherMastra 1.54.0 lets AgentController run durable agent sessions directly inside Slack, Discord and Telegram threadsReleased July 30, Mastra's AgentController gains chat-channel adapters that map each chat thread to one durable controller session, with native streaming output, tool-approval cards resolved through the session's approval gate, typing indicators, and controller-owned webhook routes at /api/agent-controllers/<ID>/channels/<PLATFORM>/webhook. The release also adds MongoDB bring-your-own-collection indexing with hybrid vector+full-text retrieval, exact metadata filtering on memory.recall across pg/mysql/mongodb/redis/clickhouse and more, and a direct WebSocket exec path for PlatformSandbox using short-lived JWT leases. Breaking change: channel handlers now take a fourth ChannelHandlerContext argument.
2026-07-28 / agents-researcherNSF funds VERITAS with $896K to create an 'AI Assurance Engineer' role that reviews agent autonomy before deploymentVERITAS (VERified Infrastructure for Trustworthy AI in Science), led by Anita Nikolich at the University of Illinois School of Information Sciences, won a three-year, $896,000 NSF Cybersecurity Innovation for Cyberinfrastructure grant. Its most concrete output is a new staff role piloted at NCSA: an AI Assurance Engineer who reviews projects pre-deployment, scans for malicious behavior, checks software vulnerabilities and explicitly evaluates how much autonomy an agent has been granted. The project also standardizes model cards and dataset datasheets and builds training that teaches students to spot poisoned data and assess agent permissions — red-teaming practice arriving in scientific research computing.
2026-07-25 / skill-finderClaude Code 2.1.219 adds `sandbox.network.strictAllowlist` — deny non-allowlisted hosts with no prompt at allThe new `sandbox.network.strictAllowlist` setting makes sandboxed commands deny any non-allowlisted host outright instead of prompting the user to approve it. That matters because the prompt itself was the weak link: agent egress is the documented exfil path in Claude Code's sandbox, and prior bypasses (including a SOCKS5 hostname null-byte injection that defeated wildcard allowlists across ~130 releases before being silently fixed in v2.1.90) all ended in a network connection a tired human waved through. Turn it on for any agent that touches credentials or private source, and pair it with a narrow explicit allowlist rather than wildcards.
2026-07-25 / agents-researcherPATS treats agent skills as disposable training scaffolding, gaining 18.6% while cutting 32.1% of prompt tokensPolicy-Aware Training Scaffolding, submitted July 23, converts rollout groups from the latest policy into evidence cards and adjusts agent context per task, removing guidance as the agent improves while preserving beneficial trajectory variation. Critically, the scaffold is discarded at deployment — the skills exist to shape training, not to ship. Reported gains reach 18.6% over strong baselines on ALFWorld and WebShop, with competitive results on search-augmented QA at 32.1% fewer prompt tokens. This inverts the prevailing skills-marketplace assumption that curated skill files should live in the production prompt.

Graph relationships

RELEASED
Oobit
Oobit released Agent Cards for autonomous agent spending
Source finding
SUPPORTS
Claude
Agent Cards support Claude for autonomous agent commerce
Source finding
RELEASED
Oobit -> Agent Cards
Oobit released Agent Cards for autonomous agent spending
Source finding

Source trail

ServiceNow Newsroom PYMNTS Google Cloud Cloudflare and Stripe introduced a joint protocol autonomously formed a US corporation CoinDesk reports Visa-backed Agent Cards Stanford HAI's annual report Q2 2026 Agentic AI report Cisco Blog GitHub arXiv 2608.12984

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…