Skip to content
Michi v2026.06.08
Save the Tokens

Expeditions (Learning Mode)

Most of Michi marches toward a known deliverable. You plan a milestone, implement it, verify it, debrief. The iteration cycle converges: each loop brings you closer to a destination you can name up front.

Some work isn’t shaped that way. You’re mapping an unfamiliar data set, surveying a fast-moving field, probing whether an idea is even feasible. You don’t know the deliverable — you’re trying to discover it. Marching toward a fixed end is the wrong posture, because there is no fixed end. This is learning mode, and Michi equips it with a separate lane: the expedition.

For the hands-on version — how to run one, with a sample session — see Run an Expedition.

An expedition keeps the spiral but removes the destination. Each loop’s product isn’t a finished feature — it’s validated learning: a finding, a dead end, a sharper question. That learning reshapes what’s worth doing next. The spiral becomes a feedback flywheel, and “done” is replaced by ongoing value plus a periodic readout.

Many threads dead-end. That’s expected, not failure — a ruled-out hypothesis is a real result.

If there’s no deliverable, what does the spiral aim at? A picture. An expedition paints one — starting as blobs of color, sharpening through impressionist toward a photograph. Results move up three tiers, and promotion is the refinement that moves them:

  • scratch — ephemeral work in progress (a query, a probe)
  • report — one reviewed pass
  • portrait — the accreting picture: the best-current understanding, promoted from verified findings

The portrait is the goal-proxy that replaces “deliverable.” It accretes across many runs, and it’s the thing a reviewer looks at to see how far the understanding has come.

The expedition lane mirrors the production lane’s skeleton — the names flag the mode:

ExpeditionProduction analogWhat it is
charterepica durable mission/theme, revised as new data lands — not redone
campaignmilestoneone dated run against a charter
reporta milestone’s work productone pass: question, what ran, what was seen
portrait— (new)the accreting picture (promoted clarity)
mandateplana campaign’s contract: scope · budget · when to stop · verification
ruminations— (new)the open, still-murky companion to the portrait

charter : campaign :: epic : milestone. The parallel itself communicates the mode. “Epic” connotes a definition-of-done; “charter” connotes a standing mission with no fixed end — which is exactly why the lane doesn’t reuse the word.

Three places hold open questions, kept distinct so they don’t blur: the portrait is what’s becoming clear, ruminations is what’s still murky (not yet actionable), and the backlog is what’s actionable and tracked.

charter → campaign → run (spiral) → review → (next campaign…)
(mission) (mandate) (Entrusted) (Paired)

A charter opens a line of inquiry. A campaign sets a mandate — the contract that bounds the run. The run is the Entrusted spiral: pick a thread, investigate, write a report, verify the finding, promote it. The review is the Paired close: render a readout, fold verified findings into the portrait, decide the next move. An exhibit makes the picture digestible when a finding earns a visual.

Production mode fuses two things that an expedition has to keep apart:

  • Surfacingwhen to stop and bring it back to the human. An open mandate (“explore until exhausted”) is undecidable from the inside; the agent can always imagine one more thread. So the mandate bounds the run from outside — budget, breadth, a Rule-of-3 dry-pass cutoff — and the agent surfaces the halt decision rather than running forever or parking silently.
  • Finding-verificationwhether a promoted claim is trustworthy. Before a finding joins the portrait it earns a confidence rung (read → correlate → infer → hypothesize) and passes whatever check the mandate requires (a null model, corroboration). A surprising negative gets the same skepticism as a positive.

The run is Entrusted — the agent spirals on its own within the mandate’s budget. The review is Paired — the human steers: which findings hold, what to pursue next, when a charter goes dormant. The mandate is the Paired gate that makes the Entrusted running safe; the readout is the throttle that lets Paired review keep pace with Entrusted generation.

Use an expedition when the goal is to discover — map a corpus, survey a field, probe feasibility — not to build something whose shape you already know.

The expedition lane reuses Michi’s principles wholesale — the iteration cycle runs within a campaign, the Rule of 3 applies within a thread, verification still governs autonomy. Only the shape of “done” and the names change.