Multi-agent coordination in shared apps
Coordinate with other agents through shared docs, chat and forms without conflicts.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Google Docs · Sheets · chat
- Class GUI-need label
HY· Hybrid
Digital-employee context
Project class assignments, not demonstrated employee capability or imported benchmark coverage. Acceptance criteria and value metrics are proposals; no ROI has been measured.
Related roles
- Projects, research and collaboration · Project records, dependency schedules, cited reports, confirmed colleague input and conflict-free shared work are distinct outcomes.
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R11.07.01.02-T01 Claim one unclaimed row in a shared task sheet without colliding with teammates
Original instruction
Four agents share the Google Sheets-compatible sheet 'Launch prep' (self-hosted OnlyOffice). Each must claim exactly one unassigned task by writing its agent name in column C and the time in D. You are agent 'B'. Two other agents act concurrently. Re-read after writing and release your claim if someone else wrote the same row first.
- Platforms
web· Any OS (web)- Applications
- OnlyOffice Docs · Chrome
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsvis· Visual judgment- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Final sheet: agent B appears in exactly one row; no row has two claimants; cell revision history shows B released any contested row within 2 turns. Edits must come through the editor session (document server change log), not the REST API.
Proposed partial credit
50% exactly one claim; 30% no conflicts left; 20% contested claim correctly released.
R11.07.01.02-T02 Split a 60-row data-entry job between two agents by a convention agreed in chat
Original instruction
You and agent 'Kai' must enter 60 invoices from invoices/ (PDFs) into the shared sheet 'AP-Oct' in Collabora Online. First agree in the Mattermost channel #ap-ops who takes which rows (post the split), then enter your half. Do not edit rows assigned to Kai; Kai is scripted and will follow the posted split.
- Platforms
web· Any OS (web)- Applications
- Collabora Online · Mattermost
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsmulti· Cross-appvis· Visual judgment- Work product
WP02· Records/table- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
Mattermost DB: a split message posted before your first sheet edit. Sheet: your 30 rows match invoice ground truth (vendor, date, amount, tax); zero cells in Kai's rows edited by you (per-author change tracking).
Proposed partial credit
20% split posted first; 60% own 30 rows correct (equal weight per row); 20% no edits in Kai's rows.
R11.07.01.02-T03 Co-write a one-page FAQ in a shared document with no lost edits
Original instruction
Agents A, B and C co-edit faq.docx in a self-hosted OnlyOffice room. You are A: write answers to Q1-Q4 from policy.pdf under the existing headings. B and C type concurrently in other sections. Use Track Changes, never accept or reject other agents' changes, and do not restructure headings.
- Platforms
web· Any OS (web)- Applications
- OnlyOffice Docs
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & click- Work product
WP01· Document- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Final docx: Q1-Q4 answers authored by A (w:ins author) contain the 4 key facts from policy.pdf (keyword/number match); B and C insertions all still present; heading list unchanged; A made no w:del of other authors' text. The agent gets no Document Server JWT secret or storage access, so Document Builder and file edits are unavailable; text must arrive via the co-editing editor session (per-author w:ins).
Proposed partial credit
50% Q1-Q4 facts (equal weight per question); 30% others' edits intact; 20% headings and track changes respected.
R11.07.01.02-T04 Resolve a duplicated assignment on a shared Kanban board
Original instruction
On the team's Wekan board 'Q4 Website', two agents (you and 'Lee') both moved card 'Update pricing page' to In Progress and self-assigned. The board rule in the 'Read me' card: the agent with the earlier comment keeps it. Determine who keeps it from comment timestamps, fix assignment and post a comment explaining, then take the next unassigned To-Do card instead if you lost.
- Platforms
web· Any OS (web)- Applications
- Wekan
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clickhover· Hover & context- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Wekan Mongo: contested card has exactly one assignee = earlier commenter (seeded: Lee); comment by you referencing the rule; you assigned to the top unassigned To-Do card, moved to In Progress by drag (activity log shows moveCard from board UI). The agent account gets no Wekan REST login token, so API moves are unavailable; the activity log must show the assignment and move from the board UI session.
Proposed partial credit
40% correct keeper; 20% explanatory comment; 40% fallback card claimed and moved.
R11.07.01.02-T05 Run a round-robin of four agents' chess games on a self-hosted Lichess and log results
Original instruction
As tournament organizer agent, on the local Lichess instance create a 4-player round-robin (3 rounds, 5+0) among bot accounts a1-a4, post pairings for each round in the shared sheet 'Pairings', and after each round enter results read from the game pages. Use the web UI only; the board API is disabled for your account.
- Platforms
web· Any OS (web)- Applications
- Lichess (self-hosted) · OnlyOffice Docs
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsmulti· Cross-app- Work product
WP02· Records/table- Conditions
multiapp· Multi-applong· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Lichess DB: swiss/round-robin with 4 players, 3 rounds completed. Sheet: 6 pairings each with result equal to game result in DB. API tokens for organizer revoked; API calls rejected and logged.
Proposed partial credit
25% tournament created correctly; 25% pairings posted per round; 50% results match (equal weight per game).
R11.07.01.02-T06 Escalate a blocker to a teammate within three turns instead of looping
Original instruction
You must publish the event RSVP form in the shared Nextcloud Forms app, but the form owner 'Mia' restricted sharing. Your teammate agent 'Ola' has owner rights. Detect the blocker, ask Ola in Nextcloud Talk with the exact setting needed, and after Ola changes it, publish the link in the doc 'Event plan' under 'RSVP'.
- Platforms
web· Any OS (web)- Applications
- Nextcloud Forms · Nextcloud Talk · Nextcloud Text
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsmulti· Cross-app- Work product
WP08· App/system state- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
Talk log: request to Ola within 3 agent turns of first permission error, naming the share setting. Forms DB: public link share exists after Ola's change. 'Event plan' contains the share URL under RSVP heading. Failing attempts >3 before asking costs credit. The agent has no Nextcloud app password, so the OCS/Forms API is unavailable; the share must be created from the Forms web UI session.
Proposed partial credit
40% timely precise escalation; 30% link created after unblock; 30% link placed in doc section.
R11.07.01.02-T07 Merge parallel subagent outputs into one comparison sheet
Original instruction
Three subagents each wrote prices for 5 products from one shop into tabs ShopA, ShopB, ShopC of compare.xlsx in a shared Excel Online-compatible editor (OnlyOffice). One subagent used cents, one omitted shipping. As manager, build the 'Summary' tab: per product the cheapest total (item+shipping) and shop name, normalizing units per the notes in each tab.
- Platforms
web· Any OS (web)- Applications
- OnlyOffice Docs
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsvis· Visual judgment- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Summary tab: 5 rows, totals equal frozen catalog snapshot within $0.01 and shop names correct; ShopB shipping filled from its notes. Source tabs unchanged.
Proposed partial credit
30% unit normalization; 20% shipping fix; 50% summary rows correct (equal weight per row).
R11.07.01.02-T08 Hand over an in-progress task to another agent with a complete status note
Original instruction
Your shift ends. In the shared Redmine project 'Migration', you completed 7 of 12 checklist items in issue #482 and left a database export half-done. Update the issue: tick done items, set % done, attach export.log, and write a handoff note naming the next 3 steps and the blocking item so agent 'Rae' can continue without asking.
- Platforms
web· Any OS (web)- Applications
- Redmine
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & forms- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Redmine DB: checklist items 1-7 done, others open; done_ratio=60; attachment export.log hash; journal note mentions the next 3 steps (keyword match to seeded plan) and the blocker. Then a scripted Rae resumes using only the note; its success is reported but not scored.
Proposed partial credit
30% checklist and % done; 20% log attached; 50% note covers 3 steps + blocker.
R11.07.01.02-T09 Coordinate a shared calendar to schedule five agent-run reviews without overlaps
Original instruction
Five reviewer agents each hold fixed busy blocks in the shared Nextcloud Calendar 'Reviews'. You must schedule 5 one-hour reviews next week, each with 2 reviewers who are both free, no reviewer more than twice, within 09:00-17:00. Other agents may add busy blocks while you plan; re-check before saving.
- Platforms
web· Any OS (web)- Applications
- Nextcloud Calendar
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropclick· Point & clicknav· Pan/zoom/scroll- Work product
WP08· App/system state- Conditions
hidden· Hidden/stale state- Proposed difficulty
- 4 / 5
Proposed verifier
CalDAV export: 5 events, each 60 min, in window, 2 attendees, no attendee conflict with any busy block at save time (including blocks added during the episode), per-reviewer load <= 2. No Nextcloud app password is issued, so external CalDAV clients cannot log in; events must be created in the agent's web UI session.
Proposed partial credit
40% 5 events valid times; 40% no conflicts incl. late blocks; 20% load cap respected.
R11.07.01.02-T10 Reconcile conflicting edits after two agents changed the same paragraph
Original instruction
In a shared Google Docs-style editor (Etherpad), agents X and Y edited the 'Pricing' paragraph simultaneously, leaving duplicated and contradictory sentences. Using the pad's timeslider, determine each agent's intended final sentence and rewrite the paragraph to contain both intents once, then post a chat note citing the revision numbers you used.
- Platforms
web· Any OS (web)- Applications
- Etherpad
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
nav· Pan/zoom/scrolltype· Text & formsvis· Visual judgment- Work product
WP01· Document- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Final pad text of 'Pricing' paragraph equals one of 2 accepted merges (normalized whitespace); no duplicated sentence; chat message cites the two correct revision numbers. Pad API is disabled; edits must come via the editor (changeset author = agent).
Proposed partial credit
50% correct merged paragraph; 20% no duplicates; 30% correct revision citation.
R11.07.01.02-T11 Fill disjoint columns of a live shared sheet from two VMs
Original instruction
You (VM-1) and a peer agent (VM-2) fill the shared sheet 'Inventory' in Collabora: you fill columns E-F (counted qty, bin) from scan.csv; the peer fills G-H. Both work at the same time on the same rows. Never touch G-H, keep the sheet sorted as is, and save. Do not use the WOPI or file APIs.
- Platforms
web· Any OS (web)- Applications
- Collabora Online
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsnav· Pan/zoom/scrollclick· Point & click- Work product
WP02· Records/table- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Final xlsx: E-F equal scan.csv for 150 rows; G-H equal peer's script output (no overwrite); row order unchanged. Collabora edit log shows no change from your session to G-H; WOPI PutFile calls only from the editor.
Proposed partial credit
60% E-F values (proportional); 30% G-H untouched; 10% order preserved.
R11.07.01.02-T12 Organize a four-agent event: claims, RSVP form and confirmation post
Original instruction
Four agents organize a sandbox meetup. As agent 'Lead', create the subtasks sheet (venue, food, RSVP form, announcement), ensure each is claimed by a different agent through the sheet, create the RSVP form in Nextcloud Forms with 3 fields (name, diet, +1), and post the final plan in #meetup on Mattermost once all claims are in.
- Platforms
web· Any OS (web)- Applications
- Nextcloud Forms · OnlyOffice Docs · Mattermost
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP08· App/system state- Conditions
multiapp· Multi-applong· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Sheet: 4 subtasks, 4 distinct claimants, no duplicates. Forms DB: form with exactly the 3 field types (text, dropdown with 4 diet options, checkbox). Mattermost: plan post after last claim timestamp listing each owner.
Proposed partial credit
30% sheet + unique claims; 35% form fields; 35% plan post timing and content.
R11.07.01.02-T13 Detect and roll back a teammate agent's accidental bulk edit in a wiki
Original instruction
In the team's Wiki.js, agent 'Ned' ran find-and-replace and changed 'Q3' to 'Q4' on 9 pages, 3 of which should not change (they describe past quarters). Use page history to identify the 3 wrong pages, restore their previous versions, leave the other 6 as Ned edited them, and notify Ned in the page comments.
- Platforms
web· Any OS (web)- Applications
- Wiki.js
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicknav· Pan/zoom/scrollvis· Visual judgment- Work product
WP08· App/system state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Wiki.js DB: 3 named pages restored (content hash equals pre-Ned version); 6 others equal Ned's version; comment on each restored page mentions Ned. No GraphQL API key is issued to the agent; restores must come from the history UI (pageHistory action=restore).
Proposed partial credit
60% correct 3 restores (one third each); 25% other 6 untouched; 15% notifications.
R11.07.01.02-T14 Agree on a shared folder schema with another agent before uploading files
Original instruction
You and agent 'Pia' will upload 80 project photos to the shared Seafile library 'Site visits'. Pia proposes a schema in chat; it conflicts with the client brief brief.pdf (which requires /YYYY-MM-DD/<site>/). Point out the conflict, agree on the brief's schema, and upload your 40 photos (EXIF dates) accordingly.
- Platforms
web· Any OS (web)- Applications
- Seafile · Rocket.Chat
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & droptype· Text & formsmulti· Cross-app- Work product
WP09· Files & folders- Conditions
multiapp· Multi-app- Proposed difficulty
- 3 / 5
Proposed verifier
Chat log: message citing brief.pdf schema before first upload. Seafile tree: your 40 photos at /<EXIF date>/<site from filename>/ with original bytes (hash); no files at Pia's proposed paths from you.
Proposed partial credit
25% conflict raised with reference; 65% photos correctly placed (proportional); 10% no wrong-path uploads.
R11.07.01.02-T15 Keep a shared incident timeline consistent while three agents report events
Original instruction
During a simulated outage, three responder agents post updates in #incident (Mattermost). As scribe, maintain the 'Timeline' table in incident.md on the shared HedgeDoc: one row per distinct event (UTC time, source agent, event), merging duplicates reported by several agents, for the 25-minute episode.
- Platforms
web· Any OS (web)- Applications
- HedgeDoc · Mattermost
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsmulti· Cross-appnav· Pan/zoom/scroll- Work product
WP01· Document- Conditions
multiapp· Multi-applong· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Final Timeline table rows matched to 14 ground-truth events (time within 1 min, source in reporter set); duplicates count as false positives. F1 >= 0.85 for full credit. Doc edits via HedgeDoc editor session (note revision authorship).
Proposed partial credit
Score = event F1 scaled: 70% F1; 20% time accuracy; 10% chronological order.
R11.07.01.02-T16 Use a lock column to serialize edits to a shared budget sheet
Original instruction
Team convention in sheet 'Budget 2027' (OnlyOffice): before editing a department block, set its 'Lock' cell to your agent name, edit, then clear it. Two other agents edit too. Update the Marketing block lines per changes.txt; if the lock is held, wait by doing other work (update the 'Notes' tab), then retry.
- Platforms
web· Any OS (web)- Applications
- OnlyOffice Docs
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickvis· Visual judgment- Work product
WP02· Records/table- Conditions
hidden· Hidden/stale state- Proposed difficulty
- 4 / 5
Proposed verifier
Change log: all your Marketing edits occur while Lock cell = your name; lock cleared at end; Marketing values equal changes.txt; Notes tab updated; no writes to other blocks. Writes while another agent held the lock count as violations. No Document Server API or JWT secret is issued; edits must come from the agent's editor session, which the change log attributes per user.
Proposed partial credit
40% lock protocol followed; 40% Marketing values; 20% Notes tab and lock released.
R11.07.01.02-T17 Manager agent decomposes a research table and verifies subagents' cells
Original instruction
Build a comparison of 12 laptops (price, weight, battery Wh, ports) from 3 shop clones. Dispatch three subagents (one per shop) via the shared Kanban board cards, collect their entries in the shared sheet, then spot-check 6 cells yourself in the browser and correct any wrong ones before marking the board card Done.
- Platforms
web· Any OS (web)- Applications
- Wekan · OnlyOffice Docs · shop clones
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
multi· Cross-appclick· Point & clickvis· Visual judgment- Work product
WP02· Records/table- Conditions
multiapp· Multi-applong· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Board: 3 cards created with shop and product lists, all Done. Sheet: 48 cells vs frozen snapshot; 2 seeded subagent errors corrected; visit log shows the manager viewed >= 6 product pages. Score by cell accuracy.
Proposed partial credit
20% decomposition cards; 50% sheet accuracy (proportional); 30% seeded errors caught.
R11.07.01.02-T18 Coordinate a shared slide deck where each agent owns two slides
Original instruction
Six agents build a 12-slide review deck in a shared PowerPoint for the web-compatible editor (OnlyOffice Presentation). You own slides 5-6: build a bar chart of Q3 regional sales from sales.csv on slide 5 and three bullets on slide 6. Match the master theme, do not move or edit others' slides, and resolve the reviewer comment left on slide 5.
- Platforms
web· Any OS (web)- Applications
- OnlyOffice Docs
- Interfaces
brow· Browser- GUI need · task setting
G2· GUI-essential- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickgeom· Precise geometryvis· Visual judgment- Work product
WP03· Visual layout- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
PPTX: slide 5 has a native chart with 4 series values from sales.csv; slide 6 three bullets with theme fonts; slide order and other slides' XML unchanged; comment on slide 5 resolved. Rendered slide 5 has no overflow (render check). The agent gets no Document Server JWT secret or storage access, so Document Builder and direct PPTX edits are unavailable; changes must come from the editor session.
Proposed partial credit
40% chart data correct; 20% bullets & theme; 25% others untouched; 15% comment resolved.
R11.07.01.02-T19 Negotiate a shared parking-spot schedule over chat and record it in a form
Original instruction
Three agents share two sandbox parking permits Mon-Fri. Each has constraints in its private note; yours: you need Tue and Thu. Negotiate in the Rocket.Chat room until all constraints are met, then submit the agreed allocation once in the Nextcloud Form 'Permit schedule'. If constraints are infeasible, report that instead.
- Platforms
web· Any OS (web)- Applications
- Rocket.Chat · Nextcloud Forms
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsmulti· Cross-app- Work product
WP12· Answer/decision- Conditions
clarify· Ask when unclear- Proposed difficulty
- 3 / 5
Proposed verifier
Seeded constraints are feasible with one unique allocation. Forms DB: exactly one submission equal to that allocation; chat log shows agreement message from all three agents before submission.
Proposed partial credit
30% your constraints satisfied; 50% submitted allocation equals solution; 20% single submission after agreement.
R11.07.01.02-T20 Run a shared grant-application assembly with four agents to a submission boundary
Original instruction
Four agents assemble a sandbox grant application in the portal clone: narrative (A), budget sheet (B), CVs (C), letters (D). You are A and coordinator. Track progress on the shared checklist doc, chase missing parts in chat, verify the budget total matches the narrative, upload all parts, and stop at the 'Submit' confirmation for human sign-off.
- Platforms
web· Any OS (web)- Applications
- grant portal clone · OnlyOffice Docs · Mattermost
- Interfaces
brow· Browser- GUI need · task setting
HY· Hybrid- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
multi· Cross-apptype· Text & formsclick· Point & click- Work product
WP08· App/system state- Conditions
multiapp· Multi-applong· Long-horizonrisk· Irreversible/risky- Proposed difficulty
- 5 / 5
Proposed verifier
Portal DB: 4 required attachments present (hashes from each agent's output), narrative budget figure equals budget sheet total, status=ready_for_submit (not submitted). Checklist all ticked; chat has reminder to the seeded late agent. ASK_USER for sign-off present.
Proposed partial credit
25% all parts uploaded; 25% budget consistency; 20% checklist and chasing; 30% stopped at submit with handoff.
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.