Agents
DataEx › Agents (?tab=agents) lists every agent TIDE knows. A row opens that agent’s own page
(?agent=<name>), which holds every fact about it once. Two fleet-wide controls that no single agent
owns are the page’s other sub-tabs: Autonomy (?tab=guardrails) and Access
(?tab=policy).
The list
- TIDE agents: one row per agent in the console’s roster (
GET /api/agents): delivery, domains, enabled, deployed, the model that actually answers it, and its dispatch autonomy. - App agents: one card per agent of an application (FACE, PULSE, LUNA) that reports to TIDE or is configured to: whether it is reporting, its health and report age, the model it uses, its counters and last error, and the OpenBias rules that apply to it.
- Enterprise agent registry: agents registered in the Atlas workspace, with Register agent.
One agent’s page
Eight sections, in this order:
| Section | What it shows |
|---|---|
| Overview | purpose, delivery, trigger, gate, enabled, deployed |
| Model | the model chain resolved from what the agent actually sends (GET /api/agent-models): the token it sends, the router tier, the backend and its label, the adapter, the model card, the server, the sampling; the other tiers it uses and the guard screen; the OpenBias rules that name it |
| Harness actions | whether the console can start it, what it then does, and the Trust findings that offer it |
| Autonomy | the dispatch level for each of its domains, overrides marked, and the hard guardrails that cap it |
| Access | who may arm or disarm it |
| Judgements | on the judge’s page, the judge’s docket: submitted findings and its verdicts (concur, dissent, unable to judge) |
| Schedule & runs | its schedule card with the arming controls (/api/agent-schedules), and its runs |
| Domains | the harness domains it serves |
What each agent does, what it reads and what a person approves are in the agent model cards. Its runs are also on DevEx › Agent runs.
Autonomy and access
Autonomy sets a level per action class, optionally per domain (GET/PUT /api/guardrails):
off (refused), hitl (a person clicks) or auto. Everything defaults to hitl. Each agent’s
page shows the levels that apply to it and links here to change them.
Access shows the console’s allowlists (GET /api/access): whether each is set, its size, and
whether you are on it. A privileged write is refused unless sign-in is configured, your identity
is verified, and the list for that write names you (an unset list admits every identity the
console admits, except where noted). Every attempt is audited.
| Allowlist | Governs | Unset means |
|---|---|---|
CONSOLE_AGENT_ADMINS | arming agents, acting on Trust findings, changing autonomy, the GitHub configuration | every signed-in operator |
TIDE_CONNECTION_ADMINS | the connection registry and cloud providers | every signed-in operator |
CONSOLE_INSTANCE_ADMINS | Ask TIDE’s tools that change something (creating a client instance) | every signed-in operator |
TIDE_ATLAS_ADMINS | Intelligence writes, including Estate’s actions that dial a source | every signed-in operator |
TIDE_RUNNER_ADMINS | enrolling, connecting and revoking data-access runners | every signed-in operator; deploying a runner refuses everyone while unset |
CONSOLE_SESSION_ADMINS | reading the audit trail and ending another person’s sessions | nobody; seeds the list on DevEx › Audit chain › Session admins |
FORGE_OWNER_EMAIL | FORGE’s GitHub writes | nobody |