Part 2 of The AI Cohesion Paradox. Part 1 named the problem. This part is about the habits and processes that fix it.

In May I argued that unmanaged AI adoption turns teams into islands. Each person, paired with an agent that never pushes back, becomes highly productive and quietly disconnected. Departments drift into “sovereign states.” The informal fabric of an organization, the hallway question and the “who knows about X?” ping, gets used up without anyone measuring it.

The question readers asked most was simple: so what do we actually do? This follow-up answers it. It is deliberately light on tools. Tools can hold shared context, but they cannot create the will to share it. That comes from habits, rituals and incentives. This is the social architecture from Part 1, made concrete.

The thesis in one line: an island is a person with an agent; an archipelago is a set of islands joined by deliberate bridges. The bridges are mostly habits, not software.

What changed since Part 1: the evidence got sharper

Six months of new data point the same way. Individual gains are real, collective gains are not automatic, and the gap between them is social.

The asking stopped, even at the source. Anthropic studied its own staff in August 2025: a survey of 132 engineers and researchers, 53 interviews and internal Claude Code usage data. Claude became the first stop for questions that used to go to colleagues. Some engineers reported fewer mentorship and collaboration opportunities; one said junior people no longer come to them with questions as often. If this happens at the company building the model, it happens everywhere.

The public commons collapsed too. Stack Overflow received 17,935 questions in December 2024 and 3,312 in December 2025, a drop of 81.5%. The decline started before ChatGPT, but AI turned a slope into a cliff. The same thing is happening inside companies. Nobody plots that chart.

Output moved the bottleneck onto people. Faros AI telemetry from more than 10,000 developers shows high-AI teams completing 21% more tasks and merging 98% more pull requests. But review time rose 91%, PRs grew 154% and bugs per developer rose 9%, while lead time and change failure rate stayed flat. Integration, the part that used to happen in conversation, became the constraint. This is Part 1’s railway problem, now measured.

AI amplifies the system it lands in. The 2025 DORA report (nearly 5,000 technology professionals, about 90% AI adoption) concludes that AI’s main role is as an amplifier. It strengthens strong organizations and exposes weak ones.

Combining humans and AI doesn’t automatically produce synergy. Vaccaro, Almaatouq and Malone analyzed 370 effect sizes from 106 experiments. On average, human–AI combinations performed worse than the best of humans or AI alone. They gained on creative tasks and lost on decision tasks. The authors’ conclusion is the one this essay builds on: what’s missing is process.

Managers are the lever, and they are tired. Gallup’s 2026 State of the Global Workplace puts global engagement at 20%, the lowest since 2020, with 22% of employees reporting daily loneliness. Employees whose manager actively supports their team’s AI use are 8.7 times as likely to say AI has transformed how work gets done. Yet fewer than one in three strongly agree their manager does this, and leaders report more loneliness than individual contributors.

Secrecy fills the gap when norms are missing. In WalkMe’s 2025 survey of 1,000 US workers, 48.8% had hidden their AI use to avoid judgment, rising to 53.4% among C-suite leaders. In an Infosys and MIT Technology Review Insights survey of more than 500 executives, 83% say psychological safety measurably increases AI success, but only 39% rate their own organization’s safety as very high.

The best adopters treat AI as a team sport. This is the most important finding of the year. Microsoft’s 2026 Work Trend Index surveyed 20,000 AI users in 10 countries and identified “Frontier Professionals,” the 16% who get the most from agents. Compared with their peers, they are:

  • far more likely to brainstorm and refine processes together with their team (63% vs. 32%)

  • far more likely to share AI tips, new agents, lessons and mistakes (61% vs. 36%)

  • more likely to discuss quality standards for AI-assisted work (54% vs. 29%)

  • more likely to pause before a task and decide whether AI or a human should do it (53% vs. 33%)

  • more likely to do some work without AI on purpose to keep their skills sharp (43% vs. 30%)

The people getting the most out of AI are the ones who made it social. Their managers are visible too: 85% openly use AI themselves, against 64% for everyone else.

What “the shared space” actually is

The shared space is the team’s memory of who knows what. Today it is quietly eroding.

Psychologists call it a transactive memory system, Daniel Wegner’s term for a group’s shared map of who knows what. Nobody holds all the knowledge. Everyone holds a pointer to the person who does. That map is built and refreshed through asking. Each question to a colleague does two things: it gets you an answer, and it updates the team’s sense of where expertise lives.

When the question goes to an agent instead, the first part still works. The second part stops. The quiet failure of AI-heavy teams is not that knowledge disappears. It’s that the team forgets who knows things, including what each person’s agent has learned in sessions nobody else saw.

This is fixable. Hopf and colleagues (Journal of Information Technology, 2025) studied hybrid teams and found that humans and intelligent agents can build a joint transactive memory. Whether people treat the agent as part of team memory or just a private notebook depends on how complex the task is and how capable the agent is. So it can be done, but only by design.

In practice, the shared space has three channels:

  1. 1.

    Human to human. Trust, judgment, weak ties, mentorship, disagreement. AI can’t replace this channel, and AI use wears it down unless someone protects it.

  2. 2.

    Human to agent. Intent, context and review. Today this channel is almost entirely private: prompts, sessions and hard-won lessons stay in one person’s chat history.

  3. 3.

    Agent to agent. Specs, contracts and shared memory. This channel barely exists yet. Agents inherit whatever humans agreed on and wrote down, nothing more.

The key point is the direction of dependency. Channel 3 can only be as good as channel 1. Agents align on what humans have agreed. If the humans haven’t talked, every agent integrates against a different guess. The shared space flows one way: human agreement, then written shared context, then agent alignment. Skip the first step and the last one fails.

Seven habits that rebuild the bridges

Habits are what people do without a meeting telling them to. These seven are small enough to start on Monday, and each one targets a specific failure from Part 1.

1. Ask humans “why,” ask agents “how.” Send questions about syntax, APIs and lookups to the agent. Send questions about intent, history, trade-offs and “why is it like this?” to a person. The “why” questions are the ones that build weak ties and keep the team’s who-knows-what map current. Frontier Professionals already work this way: they pause before a task to decide whether AI or a human should handle it.

2. Work out loud with your agent. Share sessions, not just results. When an agent run teaches you something, such as a prompt that worked, a dead end or a surprising failure, post it where the team can see it. Shopify’s AI memo put it plainly: learning is self-directed, but share what you learned, wins and losses included. This turns private work with an agent back into team knowledge.

3. Make disclosure boring. “AI-assisted” should be routine metadata, like a reviewer’s name, not a confession. Hidden use is what breeds the trust tax: workslop, suspicion and the 48.8% who hide. Leaders go first. When a manager openly says “my agent drafted this, here’s what I changed,” the norm is set.

4. Agree on intent before you start generating. Before you let agents loose on anything that touches someone else’s area, write a short intent note covering why, constraints and what done looks like, and agree it with one colleague in a ten-minute conversation. This is cheap insurance against the railway problem. Two crews whose agents start from the same agreed intent can still meet in the middle.

5. Pair across islands. Run regular sessions where two people from different functions drive one agent together. The Harvard/P&G “cybernetic teammate” experiment broke down silos because of deliberate cross-functional pairing, not because of the AI alone. Shared screens rebuild shared mental models.

6. Keep some reps unplugged. Do some work without AI on purpose, especially reviews and design. Frontier Professionals do this more than their peers (43% vs. 30%). For juniors, protect explained-by-a-human reviews: the mentorship Anthropic’s engineers said they were losing is exactly where judgment gets passed on.

7. Review as conversation, not gate. Keep AI-generated changes small. For anything large, the author walks the reviewer through the intent live. The reviewer asks “why?” rather than “what?” With PRs growing 154% in AI-heavy teams, the review is now the main place where humans still meet over the work. Treat it that way.

Processes: making habits survive the calendar

Habits fade without structure. The organization’s job is to make the right behavior the easy default. Here are six process changes, ordered from cheapest to hardest.

1. Rituals that put the hallway back.

  • A weekly agent retro, 30 minutes. Each person brings one win, one failure and one prompt, skill or instruction worth keeping. What survives goes into the team’s shared context. This is habit 2 with a calendar slot.

  • A collision check instead of a status standup. Agents can write status updates. What they can’t do is notice that two people’s agent work is converging on the same boundary. Spend the standup on one question: whose work touches whose this week?

  • Human moments at the seams. Deloitte’s “human moments” idea works best when tied to workflow stages. Kickoff and integration should always be synchronous, human-to-human conversations. Everything in between can be as agent-driven as you like.

2. A context steward role. Every team needs someone who curates its shared context: agent instructions, decision records, the glossary, definitions of done. Rotate the role monthly so it doesn’t become one person’s island. The steward doesn’t write everything. They make sure what the team agreed actually gets written down where agents and newcomers will find it.

3. Agent context as team property. Treat shared agent instructions (CLAUDE.md files, skills, specs) like code. Changes go through review and discussion, with the team owning the result. This one move turns the private human-to-agent channel into a social object people argue about. That argument is the point.

4. Plan for integration capacity. If agents raise output and review time rises 91%, then review is a capacity problem, not a discipline problem. Set limits on PR size and on unmerged agent work in flight. Count review and integration as real work in sprint planning, not as overhead squeezed in around “real” output.

5. Incentives that reward sharing. Measure team outcomes, not individual throughput, or you will pay people to stay on their islands. Shopify added AI usage to its performance and peer reviews. Go one step further and ask what each person taught or shared. In Microsoft’s data, Frontier Professionals are more than twice as likely to say they are rewarded for reinventing work with AI regardless of the outcome (26% vs. 11%).

6. Safety and stance from the top. Publish a clear, communicated AI stance, which DORA identifies as a foundational capability. Then back it with behavior: no penalty for honest disclosure, leaders sharing their own use and failures, and room to experiment. Ethan Mollick’s Leadership–Lab–Crowd model is a useful frame here. Leadership sets direction and permission, a Lab turns experiments into tested workflows, and the Crowd, your employees, finds the use cases. Without the safety, the Crowd goes underground and you get secret cyborgs instead of a shared practice.

One more protection deserves its own line: guard the junior path. Guarantee juniors a fixed number of human mentoring hours and explain-it-back reviews. A team that stops teaching juniors is borrowing against its future senior engineers.

Where tools fit (briefly)

Tools can’t create the will to share. They can lower its cost and make what’s shared durable, and readable by both humans and agents. The test for any tool: does it help the agent route people to each other, or around each other?

I’ll use my own project, Typed Graph, as an example, because it was built for exactly this gap. It turns plain Markdown notes in Obsidian into a typed, labeled property graph. A link stops saying just “related” and starts saying how: owns:: [[Billing API]], decided_by:: [[Alice]], knows:: [[Bob]] {since: 2020}. The vault can be queried with openCypher and exposed to agents through MCP and GraphRAG. Here is how that supports the habits above:

  • Transactive memory made explicit. Edges like owns:: and knows_about:: turn “who knows what” into a map you can query. An agent that can read the map can answer “talk to Dave, he owns this” instead of answering in Dave’s place. That’s a bridge, not a replacement.

  • Decisions with their reasons. because:: and supersedes:: edges let agents read why something is the way it is, not just what it is. That protects the “why” questions from being answered with confident guesses.

  • One glossary, many agents. Shared definitions of done, domain terms and boundaries live as typed nodes that every team member’s agents load. This is the written step between human agreement and agent alignment.

  • Markdown stays the source of truth. It lives in git, gets reviewed and gets argued about. The shared context remains a social object, not a black-box index.

The same logic applies to shared CLAUDE.md files, skills, spec-driven tooling and organizational network analysis. Each is useful, and each is secondary. A well-typed graph that nobody talks about is just a nicer-looking island.

How to tell it’s working

If you only measure output, you will optimize for islands. Track a small set of social signals next to your delivery metrics, as trends over time rather than targets.

  • The asking ratio. Add Part 1’s pulse question, “How often did you ask AI something you’d previously have asked a colleague?”, and pair it with its mirror: “How often did a colleague ask you something?” When the second number falls, the team’s who-knows-what map is decaying.

  • Cross-team tie density. Run an organizational network analysis baseline before the next wave of agent rollout, then repeat it every quarter.

  • Disclosure comfort. The share of people comfortable telling their manager how they use AI. It predicts adoption and protects trust.

  • Integration health. PR size, review time, rework and merge conflicts. These are where the railway problem shows up first.

  • Shared-context vitality. How many people contributed to the team’s agent instructions and decision records this month, not just how many pages exist. One contributor means it’s still an island.

  • Junior growth. Mentoring hours, and how long it takes a new hire to make their first independent design decision.

  • Workslop received. “In the last two weeks, did you receive work that felt obviously AI-generated and shifted effort to you?”

From islands to archipelago

John Donne’s “no man is an island” was a description of the human condition. In the AI era it has turned into a design requirement. Everyone now has the means to become a productive island. The organizations that pull ahead will be archipelagos: islands that stay connected by deliberate bridges, so that people, ideas and context keep moving between them.

Those bridges are mostly habits: asking a human why, sharing a session, agreeing on intent before generating, reviewing as a conversation. Process keeps them alive. Tools make them cheaper. None of it happens by default, because the friction that used to build these connections for free is exactly what AI removed.

Start small. This month, pick one habit for yourself and one ritual for your team. Measure the asking ratio before and after. Then tell me what happened. This newsletter is my own small attempt at working out loud.

The agent-to-agent channel needs its own essay. Promise Theory offers a good vocabulary for it, and that’s where this series goes next.

Sources