I have worked for leaders who held the whole picture in their head. On purpose. The roadmap lived in one notebook. The reasoning behind last quarter's architecture decision lived in one skull. You asked a question, you got an answer sized to your job title.
For a long stretch, this worked. Being the only person with the full map made you hard to replace and easy to promote. People called it decisive. People called it a safe pair of hands.
Those days are closing out, and not because leadership fashion moved on. They are closing out because the tools your team touches every day now write down what happened, who asked for it, and what the machine did next.
Your tooling files a report on every piece of work
Look at what a coding agent emits now. Claude Code ships OpenTelemetry export out of the box. Anthropic's monitoring docs list the metrics: sessions started, lines of code modified, token usage, session cost in dollars, commits created, pull requests opened, and the accept-or-reject decision on every code edit tool call.
Then the events. A structured log line for each user prompt, each model response, each API error, each API refusal, each tool result, each permission decision, each MCP server connection. Traces tie one prompt to the API requests and tool executions underneath it, correlated by a prompt ID, so a reviewer walks the whole chain of a single piece of work from ask to commit.
Now read the attribute list on all of it. Session ID. Anonymous user ID. Account UUID. Tagged account ID. Email address when authenticated. Organization ID.
The work your team does with AI lands in a collector, timestamped, attributed, and queryable by anyone with dashboard access. Prompt text and tool output stay redacted until someone sets OTEL_LOG_USER_PROMPTS=1, so the content stays under your control. The shape of the work does not. Who asked for what, how often, what got approved, what got rejected... all of it goes into a database with a name attached.
This is not surveillance creep. It is the same telemetry discipline we pointed at services fifteen years ago, now pointed at knowledge work. And knowledge work is exactly where the lone wolf used to hide.
The regulator put it in writing
Article 12 of the EU AI Act requires high-risk AI systems to "technically allow for the automatic recording of events (logs) over the lifetime of the system." Read the text of Article 12 yourself. The detailed minimum in 12(3) covers biometric systems: start and end time of each use, which reference database got checked, the input data behind a match, and identification of the people who verified the result.
Identification of the people who verified the result. A regulator wrote into law the duty to name the human who signed off.
The general logging duty in 12(1) covers every system on the high-risk list, and Annex III point 4 puts employment and workers' management on it: recruitment and selection, filtering job applications, evaluating candidates, decisions on promotion or termination, task allocation based on individual behavior, and monitoring or evaluating the performance of people at work.
So if your company runs AI anywhere near hiring, promotion, or performance, the log is no longer an engineering nicety. It is a compliance artifact. Someone outside your reporting line reads it.
DORA has been saying this for over a decade
None of this is new information about what works. Google's DORA program has been measuring it since 2014. Their finding on culture is blunt: "organizational culture that is high-trust and emphasizes information flow is predictive of software delivery performance and organizational performance in technology." Their write-up on generative organizational culture lists six behaviors of the high performers:
- High cooperation
- Messengers trained rather than punished
- Shared risks
- Bridging across silos encouraged
- Failure treated as an opportunity for inquiry
- Novelty implemented
Score a lone wolf against those six. High cooperation, no. Messengers trained, no... the messenger gets managed out. Shared risks, no, the risk sits with whoever executes. Bridging across silos, no, the silo is the whole point.
DORA also cites a 2019 State of DevOps finding: psychological safety predicts delivery performance, organizational performance, and productivity. We have had the evidence for years. What we lacked was a receipt. Telemetry is the receipt.
Three moves the hoarders will try
I have watched all three.
Move one: switch off the telemetry. Reason given is privacy. Effect is a team flying blind on cost and reliability, and a leader with no answer when finance asks for AI spend per team. This buys a quarter of cover and costs the credibility of the whole function.
Move two: own the dashboard, share the summary. Asymmetry rebuilt one layer up. The data exists, a single person reads it, everyone else gets the narrated version. This holds until one engineer builds their own panel over a weekend and the narration stops matching the numbers.
Move three: weaponize the logs. Tool rejection rate becomes a performance metric. Lines of code per session goes on a leaderboard. Congratulations, you have taught your team to game telemetry, and now your numbers describe the gaming rather than the work. Goodhart's law does not care about your intent.
Every one of those moves is a leadership decision dressed up as a tooling decision. The log did not create the problem. The log removed the hiding place.
Feedback culture is infrastructure, not a perk
Here is the part most engineering leaders skip. If your systems now generate a continuous, attributed record of how work happens, the quality of the conversation around the record becomes load-bearing. A trace with no trust around it is a weapon. A trace inside a team with real feedback habits is a debugging tool for the organization.
My own research found 99.5% of survey respondents reported one or more types of bad boss. Not a rounding error. Near-universal. I write about the fix over at Step It Up HR, where the framing is awareness, acceptance, then action, in this order. Awareness first, because most bad bosses are not villains. They are unaware, and nobody ever showed them the data.
Telemetry hands them the data. Whether it helps depends entirely on whether the surrounding culture treats a surprising number as an inquiry or as an accusation.
What to do this week
Five things, in order.
- Turn the telemetry on, and say so out loud. Announce it before enabling it. Name what gets collected, name what stays redacted, name who reads it. Silent instrumentation is how you burn trust in one afternoon.
- Give the dashboard to the team, not to yourself. Same view, same access, no summary layer. If a number embarrasses you, the number is the work item.
- Write down which metrics will never touch a performance review. Publish the list. Lines of code and prompt volume belong on it. A metric with a promotion attached stops measuring reality.
- Run one blameless review off a real trace. Pick a session where an agent went sideways and the human approved something wrong. Walk the chain in front of the team. Ask what the system made easy, not who messed up.
- Audit your own information flow. List the five decisions you made last month. For each one, name who else knew the reasoning at the time. If the answer is nobody, you found your silo, and it is you.
The uncomfortable part
The lone wolf leader was never decisive. They were leveraged. Information asymmetry is leverage, and for thirty years our industry paid for it because there was no cheap way to audit it.
There is now. It shipped in your dev tooling, it is landing in your compliance obligations, and your team has already read more of it than you think.
So here is the question worth sitting with... if someone pulled every trace, every audit log, and every approval decision from your org for the last ninety days and put them on one screen, what story would the screen tell about how you lead?
Because someone is going to pull it. The only open question is whether you get there first.