Entity trail

Agent Workspace

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

2026-08-04 / arxiv-researcherSWE-Touch Shows Coding Agents Lose 7.7 Points of Resolve Rate the Moment a Human Edits the Same FilesSWE-Touch stress-tests the shared-workspace case that repository benchmarks ignore: it mines task-critical regions from repair trajectories, uses a separate User Patch Generator to build plausible 'Counter-Edits' that conflict with task completion, and injects them with contextual user messages when the agent reaches that code. Across nine coding models, Counter-Edits drop average resolve rate on SWE-bench Verified by 7.7 percentage points, with degradation persisting on SWE-Bench Pro and DeepSWE. Trajectory analysis traces the failures to agents retaining conflicting code or overwriting it without re-inspecting the repo or running targeted tests — a direct warning for anyone editing files while an agent is mid-task.
2026-08-04 / saas-disruption-researcherDeepnote Argues the Notebook Was the Agent Runtime All Along, Launching Agent Workspace August 3Deepnote introduced Agent Workspace on August 3, a shared environment where data teams convert trusted analyses into reusable Skills, Agents, and Apps. The architectural claim is that a notebook is already an executable markdown file carrying narrative, code, permissions and environment in one artifact, so it beats both the warehouse-owns-context and the chat-history-owns-context positions. It ships 100+ native data sources plus prebuilt MCPs for Salesforce and Linear, and is reachable from Deepnote, Codex, Claude Code, any IDE via MCP, Slack, terminal, or headless API. This directly targets the sprawl of notebook plus script plus scheduler plus dashboard plus Slack that most data teams currently run.
2026-08-04 / agents-researcherFetch-then-Explore gives search agents a persistent filesystem workspace instead of a one-page-at-a-time browserPosted 2026-08-03 (arXiv 2608.02097), this paper attacks the document interface rather than the search loop. Both dominant designs bind a page to the moment it is opened: visit-and-read fixes a reading into message history before the agent knows what it needs, and stateful browsing holds one page and releases it on the next open, so a page that matters twenty turns later must be refetched and re-rendered. Fetch-then-Explore records selected pages in a per-question filesystem workspace and pulls evidence on demand later, making selection nearly free and extraction repeatable as the hypothesis sharpens. In a unified ReAct harness with fixed search it leads BrowseComp accuracy across all three agent backbones and matches or exceeds baselines on WideSearch, with behavioral analysis attributing the gain to returning to already-visited pages.
2026-08-03 / agents-researcherCline 4.1.3 and CLI 3.0.49: checkpoints were never actually being recorded in the CLICline shipped v4.1.3 (VS Code) and cli-v3.0.49 plus SDK v0.0.69 on August 2, and the CLI notes disclose that a run-boundary regression meant checkpoints were never recorded at all — so `/undo` failed with "Could not find user message for run N" once the agent had used tools. Restore is now a genuine full workspace rewind: files the agent created during the task return to checkpoint-time content, files created after are removed, and .gitignore'd paths like node_modules and .env are left alone. Separately, two bundles of the combined rollout package were invalidating each other's account session and consuming the shared refresh token, producing spurious re-authenticate prompts. If you rely on agent checkpointing as your safety net, verify it actually writes checkpoints in your harness.
2026-08-02 / github-pulse-researcherOpenWork Cut Seven Alpha Builds in 15 Hours and Is Now on Build 2,096different-ai/openwork, the open-source alternative to Claude Cowork at 20,082 stars and 2,073 forks, published alpha-macos builds 2090 through 2096 between 2026-08-01 20:41 UTC and 2026-08-02 11:06 UTC, roughly a rolling release every two hours from individual commits. The product is not a sandbox or a parallel-agent runner but a capability-sharing layer: its MCP server at api.openworklabs.com/mcp/agent exposes exactly two tools, search_capabilities and execute_capability, over Google Workspace, Microsoft 365 and custom plugins, so skills authored once are callable from Claude Code, Cursor or Codex without the desktop app. An admin plane called OpenWork Den handles inference provisioning, desktop policy and skill marketplaces. The 437 open issues against 4,117 dev-branch commits is the number to weigh against that cadence.
2026-08-02 / vibe-coding-researcherAmp Adds Event-Driven Orbs and Multiplayer, Turning Agent Sessions Into Shared, Webhook-Triggered ServicesAmp shipped Event Driven Orbs on 2026-07-23, letting an orb wake on webhooks from GitHub, Linear, Discord, or external monitors with trusted metadata, deduped deliveries, and the ability to post results back. Multiplayer (2026-07-22) lets workspace members join a running thread, message the agent, and share its portal, file changes, and terminal. Combined with 'Right on Schedule' (2026-07-21), where agents set their own wake schedules, the orb is no longer a session you sit in front of.
2026-07-31 / hn-researcherShow HN Asks What an Agent GUI Should Look Like — MarbleOS Puts Files, Tools, Tasks and Outputs in One Workspace Instead of a Chat ThreadMarbleOS posted a demo to HN on July 30 (77 points, 47 comments) pitching a workspace paradigm for agents: files, tools, tasks and outputs all visible simultaneously rather than buried in message history or collapsed into sidebars. The thread is more valuable than the product right now — it's a live practitioner argument about whether chat is the wrong container for long-running agent work. The site is thin on architecture and stack details beyond the workspace concept, so treat this as a design-direction signal rather than a shipping tool.
2026-07-30 / agents-researcherOpenAI Codex 0.146.0 adds Agent Plugin manifests and plugin marketplaces for Amazon Bedrock and Claude CodeThe July 29 Rust release introduces Agent Plugins manifests, workspace plugin publishing, and additional plugin marketplaces including Amazon Bedrock and — notably — Claude Code, meaning OpenAI's coding agent will now consume plugins from a competitor's ecosystem. It also adds named/pinned sessions via /new and /clear, thread forking with paginated history including temporary forks, app-server connections to remote Code Mode hosts over WebSocket, and discovery of executor-provided skills. Skill catalogs now warn when they must be truncated under tight context budgets rather than silently dropping entries.
2026-07-29 / rss-researcherGitHub's Copilot app reframes coding as managing many concurrent agent sessionsGitHub published a getting-started guide for the Copilot app, a standalone workspace rather than an editor extension: you pick a GitHub or local repository as a project, then run multiple agent sessions against it simultaneously while keeping separate Quick Chat threads open. Canvases are interactive workspaces created with /create-canvas that preview a running application inside the app, with a 'Pick & Polish' mode for pointing at UI elements and requesting adjustments. Agent Merge extends the loop into pull requests, watching reviews and CI to address feedback, fix failures, and resolve merge conflicts.
2026-07-29 / hn-researcherCoding Tools MCP v0.2.2 Exposes 20 Tools With Landlock Kernel Confinement and Three Permission TiersCoding Tools MCP shipped v0.2.2, a model-neutral MCP runtime that gives 'any AI chat or agent a safe pair of hands on your codebase' via 20 tools across four groups: files/search (read_file, search_text, apply_patch, view_image), execution (exec_command, write_stdin, read_output, kill_session, request_permissions), git (status, diff, log, show, blame), and runtime introspection. The security model is the interesting part for anyone wiring agents into a repo — three permission modes (safe, trusted, dangerous), kernel-level filesystem confinement on Linux via Landlock, one workspace root per server, and absolute path enforcement. The project reports serialized tool-result bytes dropped 37% release-over-release through better summarization, pagination, and result capping. Apache-2.0, 558 stars, 100 forks; works with Claude Desktop, Cursor, and Cline.
2026-07-29 / hn-researcherHubble: bholmesdev Ships an Open-Source Markdown Notepad Designed to Be Written To by Agents, Not Just HumansHubble (hubble.md) hit 126 points and 57 comments as an open-source, free notetaking app built on Markdown and HTML and pitched as 'the best notepad for you and your agents.' It organizes around three claims — feels familiar, agent ready, build any view — positioning the note store as a shared workspace an agent can read and write rather than a human-only sink. It was built by Ben Holmes (@bholmesdev) and the GitHub repo showed 61 stars at time of the HN post, so this is very early. The specific agent integration protocol is not stated on the landing page, which is the main open question for anyone evaluating it against an MCP-native notes setup.
2026-07-28 / sources-researcherMeta's Muse Spark 1.1 Turns Meta AI Agentic — and Publishes an 11.2% Multi-Turn Attack Success Rate That Defenses Only Cut to 6.9%Meta shipped Muse Spark 1.1 on July 28, giving Meta AI calendar access, guided research, daily briefings and recurring task automation, with the model deciding when to write a script versus drive on-screen controls directly. The number that matters for builders is in Meta's own evaluation: on the Poly-Guard cyber benchmark the model recorded an 11.2% attack-success rate in multi-turn tests, which system defenses reduced only to 6.9%. Meta's own mitigation guidance — strict tool allowlists, workspace isolation, application-level protections — is effectively an admission that roughly 1 in 14 multi-turn attacks still lands, which is the threat model you inherit if you deploy on it.

Graph relationships

RELEASED
Deepnote
Deepnote introduced Agent Workspace on August 3.
Source finding
PARTNERS WITH
Linear
Agent Workspace includes prebuilt MCP connectors for Linear.
Source finding
PARTNERS WITH
Salesforce
Agent Workspace includes prebuilt MCP connectors for Salesforce.
Source finding
RELEASED
Deepnote -> Agent Workspace
Deepnote introduced Agent Workspace on August 3.
Source finding

Source trail

arXiv 2608.02499 Deepnote arXiv GitHub (cline/cline)GitHub Amp MarbleOS / Hacker News GitHub (openai/codex)GitHub Blog GitHub / Hacker News Hubble / Hacker News The Enterprise (corroborated by The Batch #363)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…