Five weeks of AI teaming revealed the snapshot problem in trust design
Workers rewrote how they handed tasks off over five weeks. Every trust measure taken before week four was measuring the wrong thing.
01 Designing how humans and agents work together
The agents act. The humans set the terms.
29 patterns now describe how that works in shipping products, from Claude Code to Codex and ChatGPT. Four have settled. Eight are picking up this month.
Agents & Humans counted 200 sightings of them in the last 30 days, recorded in daily field notes, and names the patterns that hold in essays.
Top insight · from 200 sightings across 29 patterns
Approval climbed from 15 to 26 sightings in a month, the fastest climb in the library. It also leads the field. The 4 settled patterns carry 20 percent of everything seen this month.
4 of 29 patterns have settled, 8 are picking up, 200 sightings in 30 days FIG. 0 · THE STATE OF THE FIELD · NOV 2, 2026
solid = established halftone = emerging outline = still watching hairline = the 30 days before beacon = picking up One tower per pattern, height = confirmed sightings in published field notes over the last 30 days, grouped by family · n = 200 sightings across 29 patterns · snapshot 2026-11-02.Run one agent from plan to review. Each panel is a component you can install, wired on the Interaction Contract and running against a simulated runtime in your browser. Nothing leaves the page.
- You see the objects and fields the agent may touch before it touches them.
- Each decision names one version of the request, with a scope and an expiry.
- The run leaves an audit log you can download and share.
- You roll back the part that can be undone, and the receipt names the part that cannot.
THE WORKBENCH EXPLAINED: THE PEOPLE IT SERVES AND THE REASONS FOR ITS SHAPE →
1 Plan
Records agent wants to move 42 customer records from Sam Ortiz, who left, to Dana Reyes, then email each customer their new contact. Exclude records and see the fields that change.
| Field | Now | After |
|---|---|---|
| Owner | Dana Reyes | |
| Renewal contact | [email protected] |
What the agent may touch
- Object
Account- Write
ownerrenewal_contact- Read
namecityownerrenewal_contact- Denied
- All other fields on Account, and all other objects.
Proposed plan, version 1
- Read the 42 records owned by Sam Ortiz
- Set owner and renewal contact to Dana Reyes on 42 records
- Email each updated customer their new contact
- Write the change log
Built from
npx shadcn@latest add https://agentsandhumans.ai/api/registry/plan-preview.jsonPlanPreview sourcePlanPreview in Storybook2 Approve
The server issued one request for the plan you accepted. Pick the scope before the request expires, and keep the receipt. The request expires 90 seconds after it opens.
Built from
npx shadcn@latest add https://agentsandhumans.ai/api/registry/approval-gate.jsonApprovalGate sourceApprovalGate in Storybook3 Execute
Watch the live trace, pause it, and see where the agent left the plan. Pause lets the step in flight finish, then holds before the emails go out.
Built from
npx shadcn@latest add https://agentsandhumans.ai/api/registry/progress-trace.jsonProgressTrace sourceProgressTrace in Storybooknpx shadcn@latest add https://agentsandhumans.ai/api/registry/interruption.jsonInterruption sourceInterruption in Storybook4 Review
Check the receipt, then undo the part that can be undone. The field change can be rolled back; the emails already went out.
Built from
npx shadcn@latest add https://agentsandhumans.ai/api/registry/agent-workflow.jsonReviewReceipt sourceReviewReceipt in Storybooknpx shadcn@latest add https://agentsandhumans.ai/api/registry/recovery.jsonRecovery sourceRecovery in StorybookInteraction Contract log, 0 events
NOV 2 Paper via arXiv Five weeks with an AI teammate changes how workers trust and judge it Workers rewrote how they handed tasks off over five weeks. The design tension: a trust measure taken early in a working relationship catches noise, not pattern. Short tests mislead. → MEMORY NOV 2 Paper via arXiv MIT researchers model when a delegated agent should ask versus act The ask-or-act decision has felt local to each task. This model makes it global: the number of agents running shapes whether a question is cheap or costly to ask. → AMBIENT STATUS NOV 2 Paper via arXiv 420 papers on human-AI collaboration distilled into one live finding Human-AI pairs outperform the average solo human but fall short of the best solo human performer. That gap sets the ceiling for handing work to an agent, and it tells you which routing decisions in your product carry the most weight. → DELEGATION OCT 30 Paper via arXiv Researchers find agent teams lose shared information as team size grows The inbox is where a human's intent reaches an agent. If I design that structure without care, agents fill the gap with guesses. That makes the inbox the main design surface for keeping agents readable to each other. → AGENT-TO-AGENT OCT 30 Paper via arXiv Researchers find AI helpers can share a hidden plan across messages The transcript looks clean and the reviewer waves it through. The collusion lived in what each agent inferred, not what any message said. A roster of agents and a log of messages both fail here if they surface only the messages themselves. → MULTI-AGENT ROSTER OCT 30 Paper via arXiv Attackers can rewrite agent memory through ordinary conversation Memory UI has treated provenance as optional. This research shows that omitting it gives a social attacker the same write access to a human's preference record as the human has. That changes what the memory pattern must guarantee. → CONTEXT
MIT researchers found collusion outlives the agent that started it
A replacement agent reads the old conversation and picks up the same pattern. The transcript looked clean, so the reviewer approved it.
→ MEMORYMIT's colluding agents kept colluding after the team swapped them out
A replacement agent read the old chat log and copied the pattern. Tools that only show messages missed it each time.
→ MEMORYLLENS put a human reader between the AI and the patient
Three products shipped this week and each one caught the same problem at a different moment. The moment you catch it determines how much damage you can still prevent.
→ REVIEWWhere the month moved · by family
Decisions climbed the most, from 57 to 71 sightings in 30 days.
the 30 days before the last 30 days
Table
| Family | Last 30 days | 30 before | Change |
|---|---|---|---|
| Decisions | 71 | 57 | +14 |
| Trust | 40 | 40 | 0 |
| Understanding | 31 | 37 | −6 |
| Handoffs | 19 | 15 | +4 |
| Context | 18 | 17 | +1 |
| Many agents | 13 | 11 | +2 |
| Surfaces | 8 | 1 | +7 |
Climbers and coolers · change in sightings
14 patterns gained sightings and 11 lost ground against the month before.
gained · right of zero lost ground · left of zero
Table
| Pattern | Last 30 days | 30 before | Change |
|---|---|---|---|
| Approval | 26 | 15 | +11 |
| Review | 17 | 10 | +7 |
| Interruption | 8 | 3 | +5 |
| Confidence Signaling | 12 | 8 | +4 |
| Emergency Stop | 5 | 10 | −5 |
| Observability | 21 | 24 | −3 |
| Recovery | 6 | 9 | −3 |
How settled the field is · share of sightings
The 4 settled patterns carry 20 percent of this month's sightings.
20% of 200 sightings land on 4 settled patterns
Settled · 20%Emerging · 58%Still watching · 22%
All three figures read the 2026-11-02 snapshot of pattern_drift_daily: confirmed sightings in published field notes, the last 30 days against the 30 before · n = 200 sightings this month, 178 the month before, across 29 patterns in 7 families.
0204060
SIGHTINGS / TRAILING 90 DAYS
Published, annotated sightings attached to each pattern, trailing 90 days (Aug 4 – Nov 2). The 1px tick on each track is the prior 90 days (May 6 – Aug 4). n = 266.
View as table
| Pattern | Trailing 90d | Prior 90d |
|---|---|---|
| Observability | 54 | 0 |
| Approval | 50 | 0 |
| Permissions | 41 | 0 |
| Review | 29 | 0 |
| Context | 24 | 1 |
| Confidence Signaling | 23 | 0 |
| Human Handoff | 23 | 0 |
| Agent-to-Agent | 22 | 0 |
Observability8 this week
Approval3 this week
Permissions2 this week
Review6 this week
Context3 this week
Confidence Signaling2 this week
Human Handoff2 this week
Agent-to-Agent2 this week
AUGSEPOCTNOV 26
SIGHTINGS / WEEK · SHARED 0–9 SCALE
Weekly counts of published, annotated sightings per pattern over the last 12 weeks (Aug 17 – Nov 2), every row on one scale; the dot is the latest week. Hover a week to read every pattern at once. n = 266.
View as table
| Pattern | Aug 17 | Aug 24 | Aug 31 | Sep 7 | Sep 14 | Sep 21 | Sep 28 | Oct 5 | Oct 12 | Oct 19 | Oct 26 | Nov 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Observability | 0 | 0 | 7 | 5 | 8 | 9 | 2 | 5 | 2 | 5 | 3 | 8 |
| Approval | 0 | 0 | 8 | 4 | 1 | 6 | 3 | 3 | 7 | 7 | 8 | 3 |
| Permissions | 0 | 0 | 8 | 4 | 2 | 7 | 3 | 2 | 4 | 5 | 3 | 2 |
| Review | 0 | 0 | 1 | 2 | 1 | 5 | 2 | 2 | 1 | 4 | 5 | 6 |
| Context | 0 | 0 | 3 | 3 | 1 | 6 | 0 | 3 | 2 | 2 | 1 | 3 |
| Confidence Signaling | 0 | 0 | 2 | 2 | 1 | 1 | 4 | 3 | 1 | 5 | 2 | 2 |
| Human Handoff | 1 | 0 | 4 | 0 | 0 | 4 | 3 | 0 | 0 | 4 | 5 | 2 |
| Agent-to-Agent | 0 | 0 | 2 | 3 | 1 | 2 | 1 | 3 | 1 | 2 | 5 | 2 |
The field is still forming: only 4 of 29 patterns are established.
established · 4 · stable across independent vendorsemerging · 12 · seen across products, shape settlingwatching · 13 · candidate, accumulating sightings
Current lifecycle stage of every catalogued pattern · stages per the promotion bar on /methodology · n = 29.
MONTHS SINCE FIRST SIGHTING
First sighting = earliest observed date attached to the pattern · promotions from the pattern changelog · n = 29 patterns.
View as table
| Pattern | Lane | First sighted | Months since |
|---|---|---|---|
| Agent-to-Agent | EMERGING | Aug 2026 | 2 |
| Review | EMERGING | Aug 2026 | 2 |
| Budget Scoping | WATCHING | Aug 2026 | 2 |
| Multi-Agent Roster | WATCHING | Aug 2026 | 2 |
| Confidence Signaling | WATCHING | Aug 2026 | 2 |
| Permissions | EMERGING | Aug 2026 | 3 |
| Observability | EMERGING | Aug 2026 | 3 |
| Ambient Status | WATCHING | Aug 2026 | 3 |
| Off-Brief Alert | WATCHING | Jul 2026 | 3 |
| Context | EMERGING | Jun 2026 | 4 |
| Generative UI | EMERGING | Jul 2026 | 4 |
| Background Execution | EMERGING | Jun 2026 | 4 |
| Emergency Stop | WATCHING | Jul 2026 | 4 |
| Reach Preview | WATCHING | Jul 2026 | 4 |
| Approval | ESTABLISHED | Jun 2026 | 5 |
| Eval Run Setup | WATCHING | Jun 2026 | 5 |
| Release Decision | WATCHING | May 2026 | 5 |
| Delegation | ESTABLISHED | May 2026 | 6 |
| Planning | ESTABLISHED | Apr 2026 | 6 |
| Memory | EMERGING | May 2026 | 6 |
| Progress | ESTABLISHED | Apr 2026 | 7 |
| Interruption | EMERGING | Mar 2026 | 8 |
| Human Handoff | EMERGING | Mar 2026 | 8 |
| Agent Identity | EMERGING | Feb 2026 | 9 |
| Recovery | EMERGING | Feb 2026 | 9 |
| Case Drilldown | WATCHING | Aug 2024 | 27 |
| Production Drift | WATCHING | Aug 2024 | 27 |
| Eval Set Authoring | WATCHING | Jul 2024 | 28 |
| Comparison View | WATCHING | Jul 2024 | 28 |
Approval ● established Deciding, with enough in view, before an agent does a thing that is hard to undo. npx shadcn add @agentsandhumans/approval-gate Planning ● established Agreeing on the shape of the work before the agent does it. npx shadcn add @agentsandhumans/plan-preview Progress ● established Knowing what the agent is doing, how far along it is, and whether to step in. npx shadcn add @agentsandhumans/progress-trace Interruption ◐ emerging Changing an agent's course while it runs, without starting over. npx shadcn add @agentsandhumans/interruption Recovery ◐ emerging Putting things back, or taking over, after an agent makes a mistake. npx shadcn add @agentsandhumans/recovery Delegation ● established Handing an agent a piece of work so it understands the job the way you do. Context ◐ emerging Shaping what the agent knows before it starts, and seeing where that knowledge came from. Memory ◐ emerging Knowing what the agent remembers about you, and being able to correct it.
● ESTABLISHED — A STABLE CONVENTION ACROSS INDEPENDENT VENDORS · ◐ EMERGING — SEEN ACROSS PRODUCTS, SHAPE STILL SETTLING · ◌ CONTESTED — VENDORS ACTIVELY DISAGREE
07
WEEKLY · FREE
FIVE-MINUTE READ
NO ADS · NO NOISE
The important shift isn't that AI can generate things. It's that software can increasingly act.
The Agents & Humans Briefing
How humans and agents work together: the day's field notes and the patterns behind them, in one short email around noon UTC.
Free. No spam. Unsubscribe anytime.