The question that exposed an agent governance gap in almost every team we met
We build AiKey, an AI credential and governance layer for teams running models, tools, and agents in production.
Recently, one question kept creating the same awkward pause in conversations:
How many agents are running in your company right now?
Not prototypes. Not experiments from a hack week. The agents that are actually connected to real systems, touching real data, and doing real work.
A surprising number of teams could not answer it cleanly.
They knew which models they liked. They knew which demos looked impressive. They sometimes knew which team started the project.
But they could not produce a current list of active agents, who owned each one, what credentials they were using, what tools they could touch, or what spend they were driving.
That was a useful moment for us.
It made the governance gap feel much more concrete than any abstract discussion about AI safety or agentic workflows.
Agents are starting to look less like features and more like operational actors
A normal software feature does not wake up one morning with broader permissions, a new tool connection, and a higher monthly bill.
An agent can.
Once an agent is connected to email, code, CRM, docs, ticketing, or internal systems, it stops being just another model wrapper. It becomes an operating component with identity, permissions, cost, and failure modes.
That changes the management problem.
You do not just need prompts and routing. You need a current record of what exists and how it is allowed to behave.
Call it an agent register, inventory, ledger, or registry. The label matters less than the discipline behind it.
The teams that felt most in control could answer six questions fast
The strongest operators were usually able to answer six things without hunting across five tools:
- What agents are running right now?
- Who owns each one?
- Which tools or systems can each agent access?
- Which credentials or virtual keys does each agent use?
- What usage and spend belong to that agent?
- How do we revoke, pause, or investigate it if something goes wrong?
That list is not glamorous. It is also where real control starts.
Without it, every later conversation becomes slower and fuzzier.
Security asks who had access. Finance asks where the spend came from. Operations asks what should be shut off first. Nobody wants those answers to depend on a stale spreadsheet and one engineer's memory.
The tricky part is not writing the register once. It is keeping it alive.
This is where many teams break.
They agree they need better visibility, so they create a document. For two weeks it feels responsible. Then a model changes, a tool is added, an owner moves teams, a credential is rotated, and the document starts drifting away from reality.
At that point the register becomes theater.
The version that actually works has to grow from runtime activity:
- agent identity and ownership at onboarding
- credential issuance and revocation
- tool permissions and policy changes
- usage, cost, and audit events as requests happen
If the record is generated from those events, it stays useful. If it depends on manual cleanup, it dies.
This changed how we think about the product layer
One thing we have learned while building AiKey is that many companies do not need more agent demos first.
They need legibility.
They need to see which agents exist, what each one is allowed to do, what credentials are in play, how spending is attributed, and whether an incident can be reconstructed later.
That is why we keep coming back to credentials, policy, auditability, and cost ownership. Those pieces sound less flashy than agent autonomy. They are also what make agent adoption survivable once a company has more than a handful of workflows in motion.
What I would tell other builders
If you are building in the agent infrastructure space, I would not assume the next painful problem is better orchestration.
For a lot of teams, the more immediate problem is simply that their agents are no longer legible.
They have agents running, but not a trustworthy register.
They have credentials in use, but not clear ownership.
They have spend, but not attribution.
That is a very practical gap, and practical gaps usually turn into real buying urgency faster than abstract architecture debates.
That is the lens we are leaning into at AiKey: governed credentials, runtime policy, agent visibility, auditability, and cost ownership in one layer.
If your team is running into the same problem, you can learn more at https://aikeylabs.com/zh/i/ih32. For business inquiries, feel free to reach out at aikeyfounder@gmail.com.
One thing I keep noticing is that teams often have policies for what agents can do, but much less clarity around proving what they actually did.
I’d be curious whether your governance gap was more about permissions, observability, or accountability after an agent action.
The “how many agents are running right now?” question is a strong way to expose the problem. It turns something abstract into something a team can immediately recognize.