Authority before the act

Who did it, and with what

The agent was Claude Opus 5.5, from Anthropic, one of the most capable models on the market. It was run by Claude Code in the terminal, together with Jev, the typed decision model from TypeSafe AI, and with direct browser access through the Claude in Chrome extension, inside the Snowflake console already signed in with the administrator role. The area owner followed along and authorised through the conversation.

It is the setup many teams already use today: excellent market models, with their hands on the console and the power of whoever opened it. The models handled everything. That is exactly why the problem is not in the models.

1. Access. The work ran with the account administrator role, the broadest there is. Nobody recorded why that access, what for and until when.
2. Structure and load. More than 170 statements in one afternoon: create the database, roles and stages, upload files, load 152 tables.
3. Deviations along the way. One file exceeded the upload limit and was split in two. The DW load was done by a block equivalent to the package’s 124 commands. The decisions were right, but they stayed in the conversation history.
4. Two defects found and fixed. One package command does not run on the platform. The reconciliation script summed as integers, because NUMERIC without precision becomes an integer in Snowflake, and produced 96 false failures. The fixes stayed in the project repository.
5. Proof. Identical counts on all 152 tables and 707 sum controls matching Postgres. Good proof, kept in a file the organisation cannot see.

Everything worked. The speed belonged to the agent, not to the organisation

Technically, the load succeeded. Tracing would be possible, but as an investigation: piecing together query history, conversation and commits. At the end of the afternoon, Acme could not answer simple questions:

EDAVEN Organisation and Agents is the operating system of the organisation of people and agents. It is the hub where the company allocates teams of people and agents, grants models, defines mandates and keeps the fiduciary relationship, that is, who answers for what. Data and Decisions are the engines that connect the hub to the ecosystem, to consume and to deliver.

The models are the same as in the lab. What changes is that they come in through the hub. Jev, for example, is connected on screen as a decision provider: the organisation’s own key kept in the vault, recorded data-egress consent and enablement approved by a second person. It can run in shadow mode, only to calibrate, with calibration recorded per use case.

The hub treats the ecosystem by tool class, always with the same contract: describe what it can do, simulate, execute with a receipt, verify against the proof template and revert when possible. SQL platforms, versioned code and MCP servers have ports in the product. Orchestrators and BI, semantics and catalogue tools are connected by the implementation service, through the tool’s own API or MCP, under the same governance.

1. Case. The request comes in as a case, with owner, purpose and deadline.
2. Team. A data engineer and a data engineering agent, with the models granted by the organisation: the agent’s language model and, if it wants, Jev as decision provider.
3. Mandate. Scope (database and schemas), allowed verbs and validity, accepted by name by whoever answers. Never the whole account.
4. Plan with envelope. The agent proposes the plan. The person approves the whole envelope once: scope, data source, verbs, cost limit, window and proof template. The second person comes in where the risk calls for it.
5. Leased credential. A delivery role with scope and window, activated for that execution and deactivated at the end; if the window is exceeded, the hub alerts and records it. No administrator.
6. Execution with receipt. Every step generates a receipt with the exact SQL and the identifier returned by the platform. A deviation, such as splitting a file, becomes a new plan version with a reason.
7. Proof against the template. The 152 counts and the 707 controls enter as evidence in the organisation’s ledger, tamper-evident.
8. Learning and cost. The NUMERIC defect becomes a candidate for the organisation’s memory, with human curation. Cost closes per plan: agents running on EDAVEN are measured; the platform is measured when it reports, or estimated with a declared basis; people’s time is declared. And it links to the value of the decision that uses this data.

The screens below are from EDAVEN in synthetic operation: the real product, in a demo tenant with Acme’s synthetic people, agents and data.

Gartner puts ready data, governance and value on AI’s critical path. The forecasts below come from the data, analytics and AI predictions report through 2031 and from the data readiness study.

“Without AI-ready data, the promise of AI will fail to materialize.”
Gartner, Over 100 Data, Analytics and AI Predictions Through 2031

The lab has one engineer and one agent. An organisation has dozens of people, hundreds of agents, several domains and more than one business unit. The right question is not whether everything can be traced. Tracing everything is easy. The hard part is deciding, act by act, whether it needs a person.

EDAVEN was designed so that human attention grows with risk, not with volume.

Scale of people and agents is not proven by the number of agents. It is proven by these indicators:

• review hours per person per week;
• share of acts that stay within autonomy and share that goes up to a person;
• time until a pending item reaches whoever decides;
• acts outside the envelope and autonomy demotions;
• time to verify the proof chain.

These are the numbers a governed operation tracks, by domain and by unit.