Design principles

Projektor combines AI-native design with tried-and-tested principles. This page presents the evidence: the two bodies of research the design draws on, and the principles that fall out where they agree.

Two literatures matter. One is decades old and studies how humans move work through a tracker. The other is a few years old and studies how agents fail. Strikingly, they often reach the same conclusion from opposite directions.

What the flow literature says

The strongest empirical work on project management is lean/flow research, not tool-vendor guidance:

• Small batches win. The Accelerate/DORA research (Forsgren, Humble, Kim) links small batch sizes, short lead times, and frequent integration to organizational performance — the most rigorous evidence in the field.
• There is no magic WIP number. An empirical study of WIP in kanban teams found lower work-in-progress correlates with shorter lead time, but no published evidence supports a universal optimal limit. The honest advice: measure your own flow first, then set limits.
• Pull beats push. Practitioner consensus on top of the research: explicit workflow states with entry/exit criteria, a single prioritized backlog, and workers pulling the next item instead of having it pushed onto them.

What the agent-engineering literature says

• Boundaries, not vibes. Anthropic’s engineering posts — Building effective agents, the multi-agent research system, context engineering — converge on giving each agent explicit objectives, output formats, and task boundaries. Their early orchestrators spawned fifty subagents for simple queries; the fix was structural, not exhortative.
• The ticket is the dispatch unit. The emerging trackers-as-agent-infrastructure argument: API-first trackers win because the ticket carries the spec, the audit trail, and the coordination state.
• Checkpoints are architecture. Governance work (InfoWorld, enterprise patterns) puts human-in-the-loop gates at defined workflow stages and enforces policy at the point of action — not in the operator’s memory.

The six principles

Where the two literatures agree, Projektor treats the conclusion as a design constraint.

1. Tickets are self-contained specs

Humans tolerate tribal context; agents don’t. DORA’s small-batch finding and Anthropic’s task-boundary finding are the same requirement seen from two directions: work items must be small, unambiguous, and independently verifiable. A ticket ready for an agent states its acceptance criteria, names its files and scope, and includes a verification command.

2. WIP limits are enforced by the tool

For humans, WIP discipline is culture. For agents it must be mechanism, because agents make WIP inflation free — spawning ten workers costs one prompt. The fifty-subagent failure mode is what unenforced WIP looks like at agent speed. Concurrency caps belong in the tracker’s claim path, not in the operator’s restraint.

3. The tracker is the coordination substrate

Agent sessions, file claims, issue leases, heartbeats, channel messages: coordination state lives in the same durable store as the work itself, not in ad-hoc side channels. The issue graph is the shared memory; status changes are the events. (This is the thesis of Agentic workflows.)

4. One canonical workflow spec, looked up — not copied

Workflow rules drift the moment they exist in two places. The rules get exactly one home — this documentation site — and every consumer (the MCP server’s instructions, contributor docs, skills, spawn prompts) points at it rather than restating it. The server serves the spec so the deployed version and its rules cannot diverge.

5. Measure flow before tuning it

The WIP research is blunt: nobody can tell you your optimal limit. So the tracker stamps transition timestamps and exposes lead time, cycle time, and WIP-over-time — the measurements a team needs before tuning defaults like the agent concurrency cap (currently a flat default of 3) against its own baseline, not folklore.

6. Evidence is checked, not just collected

Agents pull work autonomously, and close it autonomously too: there’s no pre-close block waiting on a human. What replaces the gate is a check on the way out. Every agent-closed issue must carry a completion report (what changed, how it was verified), and that verification text is classified: a resolvable link — a PR, a CI run, a commit — passes silently; plain prose like “manually tested” gets flagged needsAudit: true for a human to pull later. The discipline isn’t blocking the transition; it’s making sure a transition without evidence doesn’t disappear into the timeline unmarked.

What this deliberately isn’t

Projektor defers the fashionable automations — triage bots, grooming bots, status-rollup agents — until the measurement and gating above exist. The literature suggests them as low-risk starting points; the same literature says you can’t evaluate them without flow metrics. Measure first.

For how these principles play out in practice — the multi-agent loop, coordination primitives, and the tool catalog — see Agentic workflows.