Skip to content

First-agent kickoff pattern

The pattern

The most successful agent kickoff (Jordan's 2026-04-14 onboarding) followed a specific multi-phase arc that should be the template for every future agent promotion:

Research session → Plan mode → Fresh-context drafting session →
Status report → Memory triage → Cutover QA → Promotion

Each phase has a natural stopping point where either Rick or chief.staff can checkpoint and redirect before the next phase proceeds.

Why it matters

Jordan's kickoff produced: - A schema design that caught a real governance contradiction in CLAUDE.md Section 10A.8 vs. 10A.9 - A correct Option D recommendation that chief.staff initially missed - A clean migration of 43 memory files with proper provenance tracking - A smoke test with HIGH confidence on the first run - Memory hygiene with clean triage dispositions - Zero defects filed at cutover

The phased structure made this possible because: - Research phase prevented drafting-with-stale-assumptions - Plan mode forced explicit A/B/C/D option comparison before commitment - Fresh drafting session kept ~30k tokens of plan-mode deliberation out of the drafting context - Status report + memory triage as separate passes prevented polish-conflation with execution

Shortcutting any phase produces measurably worse outputs.

When to apply it

  • Every agent kickoff where the agent inherits production memory (migration + triage benefits from the phased structure)
  • Every greenfield agent where the first deliverable is non-trivial (≥1 hour of work)
  • Every agent that will hold a load-bearing governance role (chief.staff, project.manager, cdo, etc.)

When NOT to apply it

  • Simple tool kickoffs where the agent's first task is mechanical (update a config, run a migration script)
  • Emergency incident response where structured phases would delay a production fix
  • Temporary agents being spun up for a one-off task with no ongoing role

Phase details

Phase 1 — Research session

Agent reads CLAUDE.md, alignment docs, their own contract, relevant memory files, and any task-specific briefs in strict sequence. Output: research notes file at /Claude/sessions/{agent}/research-notes-{date}.md capturing working hypothesis, key findings, decisions already made, open questions.

Time budget: 60-90 min for agents with substantial reading load (Jordan's was ~125 min with 43 memory files + 3 archived session exports).

Exit criterion: Research notes file committed + explicit checkpoint with Rick or chief.staff.

Phase 2 — Plan mode (if the task has an architectural fork)

Agent enters plan mode and compares options structurally (A/B/C, or more). Each option gets pros/cons, tradeoffs, implementation cost, rollback plan. Output: plan file that explicitly names the decision being made.

Jordan's plan mode caught the Section 10A.8/10A.9 contradiction specifically because plan mode forced the options to be compared side-by-side instead of defended individually.

Exit criterion: Decision approved by Rick. Agent exits plan mode with clear direction.

Phase 3 — Fresh drafting session

Agent ends the research/plan session and starts a fresh Claude Code session from the same cwd. Fresh context means the drafting isn't polluted by plan-mode deliberation or research deep-dives. The agent reads its own research notes + plan output as input, then drafts the deliverable cleanly.

This phase is critical and often skipped. Jordan's schema design v1 benefited directly from the context reset — the drafting session didn't carry the A/B/C debate forward, just the Option D decision.

Exit criterion: Deliverable file committed.

Phase 4 — Status report

Agent files a Section 6B Status Report per CLAUDE.md Section 7 universal schema. Output: /Claude/operations/status/{agent}-kickoff-{deliverable}-{date}.md.

Exit criterion: Status report filed with clear next-action language.

Phase 5 — Memory triage

Agent reviews their migrated (or newly-accumulated) memory files and classifies each as Keep / Update / Archive / Review. Output: inline triage notes in the status report OR a separate housekeeping pass file.

Jordan's triage: 4 Updates, 4 Archives, 6 Reviews deferred to Sprint 2 first-week pass. Clean dispositions.

Exit criterion: Memory triage decisions recorded. Archives executed. Updates queued.

Phase 6 — Cutover QA report

chief.staff files a Section 6D QA Report at /Claude/operations/reports/cutovers/{agent}-cutover-{date}.md with Validator Type: Scrutiny, Confidence: MEDIUM (pre-session) or HIGH (post-verification).

Exit criterion: QA Report filed.

Phase 7 — Promotion

chief.staff flips the agent from 🟡 Drafted → 🟢 Active in /Claude/system/contracts/INDEX.md. The agent is now eligible for Mission worker/validator roles.

Exit criterion: Registry updated.

Examples

Jordan's 2026-04-14 kickoff — the canonical instance. All 7 phases executed in sequence. Resulted in: - Schema design v1 approved by Rick without iteration - Option D decision that chief.staff adopted after Jordan proposed it - Clean smoke test HIGH confidence on Monday.com - Zero cutover defects - Memory triage finished same day

Contrast: chief.staff's own kickoff (retrospective) — did NOT follow this pattern because chief.staff was the first Layer 1 agent and there was no predecessor to model from. chief.staff's memory namespace is still at a legacy path (DR-002 exception) and has never had a formal cutover QA report filed. This is acknowledged technical debt.

  • [[project.manager/kickoff-package]] — the kickoff package that produced Jordan's arc
  • [[cdo/kickoff-package]], [[sop.manager/kickoff-package]], [[qa.testing/kickoff-package]] — packages following this pattern for the next 3 agents
  • [[DR-002-per-agent-cwd-memory-isolation]] — the cwd decision that makes Phase 3 (fresh session) meaningful

Source

Jordan's (project.manager) kickoff on 2026-04-14. First agent-kickoff arc executed under v1.2. Produced the pattern.

Status

Strong recommendation until formalized into SOP-EXEC-agent-onboarding-v1.0 by sop.manager (after sop.manager's own kickoff completes). At that point, this best practice becomes source material for the SOP and may be promoted to near-rule strength.