Ways of working: humans + agents

In Kera, people and AI agents share one workspace and one set of tools. You decide who drives which work — and because every action is recorded the same way, the two never drift out of sync.

Three ways in

Work enters Kera through three surfaces, all of which land in the same audited history:

Type — the app, the command palette, or the chat.
Voice — dictate a ticket or ask for a status update; transcription runs through your configured provider.
Agents — anything a person can do, an agent can do over the MCP surface. No integration to build; the capability is already there.

Bring context, not paraphrase

Attach files and images directly to tickets and documents. Give an agent the actual artifact — a screenshot, a spec, a PDF — instead of a description of it, and the context travels with the work rather than being lost in a chat.

A working loop

The everyday pattern looks like this:

Describe intent. "Stand up the Spain launch: workstreams for eng, sales, and legal, with the obvious dependencies."
An agent scaffolds structure. It creates the project, the tickets, their relations, and assignments — in one flow.
You review in the app. The result is ordinary Kera objects: open them, edit them, reassign them.
Work proceeds — by people, by agents, or both.
Ask "are we on track?" The answer is grounded in the audit trail and cycle metrics — what actually happened — not a guess.

Who does what

Agent as scaffolder: Turn a paragraph of intent into a project skeleton: epics, stories, dependencies, owners.
Agent as reporter: Answer status questions from the audit log and cycle metrics — grounded, not invented.
Agent as executor: Triage inbound work, bulk-transition tickets, keep labels and assignments tidy.
Human as decider: Approve, reprioritise, and make the calls that need judgment. The app is where you stay in control.

Kept honest

Every agent action is audited identically to a human's — same event log, same timestamps, same "who did what."
Groups gate access for people and agents alike. An agent can only touch what its identity is allowed to touch.
Identity always comes from the authenticated session, never from a request body — an agent can't claim to be someone else.

// An agent scaffolds work, then reports on it — same tools a human's UI uses create_ticket: { "project_id": "...", "type": "story", "title": "Wire SAML endpoint",
"assignee": "dev@acme.eu", "labels": ["backend"], "relations": [{ "type": "blocked_by", "target_ticket_id": "..." }] } // Later: "are we on track?" — answered from the audit trail, not a guess search_audit_events: { "resource_type": "ticket", "project_id": "..." } get_cycle_metrics: { "cycle_id": "..." }