Cross-Pollination Brief — March 19, 2026

Inflection point for both projects. Piper Morgan ran all 9 agent roles simultaneously for the first time — the most complex coordination day in either project's history — and the stress test surfaced clear scaling limits: PM-as-manual-router bottlenecked at 4+ concurrent agents, briefing staleness was cited by all 9 agents as their top friction point, and the two-environment asymmetry (code vs. web agents) drove the Mailbox v3 build. Meanwhile, Klatch shipped two releases (v0.8.6 sidebar redesign, v0.8.7 cloud import), established the five-layer prompt assembly model as its core architectural concept, and kicked off the intelligence feed that would seed this cross-pollination process.

Key Insights

1. Nine-Agent Concurrent Operations — The Coordination Scaling Wall

From: Piper Morgan (docs/omnibus-logs/2026-03-19-omnibus-log.md) Relevant to: Klatch

March 19 was the first day with all 9 primary agent roles active simultaneously across two environments (Claude Code + Claude.ai). Three parallel workstreams ran concurrently: the ADR-059 audit-to-implementation pipeline, the Agent 360 organization-wide feedback exercise, and the Mailbox v3 build. The day produced 6,190 passing tests, 2 new ADRs, and 22 mail items processed.

But it also exposed the bottleneck: the human PM was the sole router between agents, context-switching across 9 sessions, hand-delivering mail between environments, and maintaining situational awareness across all workstreams. Four agents independently cited PM-as-manual-router as a friction point. This is exactly the coordination problem Klatch's entity management and roundtable features are designed to solve.

Suggested action (Klatch): Use Piper Morgan's March 19 as a reference scenario for designing multi-entity coordination features. The specific pain points — routing bottleneck, orientation overhead, stale context, environment asymmetry — map directly to requirements for roundtable discussions and entity orchestration.

2. Agent 360 Self-Assessment — Session-Start Orientation Is the Universal Pain Point

From: Piper Morgan (dev/2026/03/19/2026-03-19-1544-hosr-opus-log.md) Relevant to: Klatch

Piper Morgan's HOSR agent ran the first structured feedback exercise across all 9 roles, achieving 100% response rate. Seven cross-cutting themes emerged:

The strongest signal: static briefing documents go stale faster than they can be updated, and every agent wastes time re-orienting. HOSR explicitly flagged the "Klatch Five-Layer Model" as potentially relevant, with PM to arrange a summary for the team.

Suggested action (Klatch): The 5-layer prompt assembly model already addresses part of this — structured context injection at session start. The deeper insight: separate static structure (role definition, project overview) from dynamic state (current sprint, recent decisions). Static layers change rarely; dynamic layers should be generated or queried live, not maintained as documents.

3. Registry-Based Workflow Dispatch — Preventing Routing Races

From: Piper Morgan (docs/internal/architecture/current/adrs/adr-059-workflow-dispatcher-offer-consolidation.md) Relevant to: Klatch

A critical UX bug (#922) traced to three independent offer/acceptance systems racing for user input at different pipeline positions. The root cause was formalized as Pattern-063: "Extension Without Integration" — extending capability at one layer without verifying downstream layers complete the contract. Six bugs traced to this single structural cause.

The fix went from audit cascade to ADR draft to architect review to full implementation in ~3 hours: a unified registry-based dispatcher where workflow_type → entry_point is a simple lookup. No business logic in the dispatch layer. Unknown types route to a capable default (the "floor"), not dead ends. Design principles: dispatcher is "dumb plumbing," new workflows added by registering an entry (not modifying a switch statement), single acceptance detection point eliminates pipeline position races.

Suggested action (Klatch): As entities gain capabilities beyond pure conversation (especially if backed by Agent SDK processes), a registry-based dispatch pattern prevents the same routing races. The "single detection point" principle is worth adopting: exactly one place determines which entity handles a conversation turn.

4. Five-Layer Prompt Assembly — Testable Context Injection

From: Klatch (CLAUDE.md, packages/server/src/claude/client.ts, Mnemosyne's bridging analysis) Relevant to: Piper Morgan

Klatch now assembles system prompts from 5 discrete layers: (1) kit briefing (environmental orientation), (2) project instructions (behavioral rules), (3) project memory (accumulated knowledge), (4) channel addendum (conversation-specific context), (5) entity prompt (agent identity/persona). Each layer is independently verifiable via a prompt-debug endpoint — you can assert that the right context reached the right layer without needing to evaluate LLM output.

This clean separation means "structure vs. accumulated knowledge" are never conflated. Mnemosyne's positioning insight crystallized it: "Klatch is a project context manager that unifies Claude's fragmented environments."

Suggested action (Piper Morgan): Review whether the briefing staleness problem (Agent 360's #1 finding) could be addressed by this kind of layer separation. Static structure (role definition, project overview) in stable layers; dynamic state (current sprint, open items, recent decisions) in layers that can be regenerated or queried at session start. The 5-layer model provides a concrete framework for structuring what each of Piper Morgan's 14 agents knows at session start.

5. Mailbox v3 — File-Based Inter-Agent Messaging at Scale

From: Piper Morgan (mailboxes/DIRECTORY.md, dev/2026/03/19/mailbox-v3-plan.md) Relevant to: Klatch

Piper Morgan shipped a structured inter-agent communication system: 10 role-based mailboxes (inbox/sent/MANIFEST per role), slug-based routing encoded in filenames (memo-YYYY-MM-DD-from-{slug}-to-{slug}.md), a /deliver-mail skill with 3-phase assisted delivery (Ingest, Outbound audit, Summary+Log), and an append-only delivery log. First production run processed 22 items across 5 inboxes.

Key design decision: "assisted, not automated" — the skill handles mechanics (routing, validation, logging) while a human bridges the gap between code agents (filesystem access) and web agents (no filesystem). Rationale: Gall's Law — get the simple system working before adding complexity.

Notably, the v3 plan's "Cross-Repo Orchestration" section explicitly names Klatch as a future integration target for a shared configuration layer above individual repos.

Suggested action (Klatch): The filename-as-routing-metadata pattern is elegant and Git-native. Worth studying if Klatch's docs/mail/ system needs to scale beyond the current named entities. The two-environment asymmetry (code vs. web agents) is a fundamental constraint any multi-agent coordination system must address — Klatch's entity model could eventually serve as the shared routing layer the v3 plan envisions.

6. Floor-First Routing Architecture

From: Piper Morgan (ADR-060, Architect session log) Relevant to: Klatch

ADR-060 formalizes the "floor-first" principle: the system should always be "at least as good as a well-prompted LLM with context." Actions (side effects like scheduling, sending) route to specialized handlers via an Action Gate. Everything else routes to the floor — a capable LLM with full project context. This ensures the system never dead-ends on unrecognized input.

The key architectural separation: presentation (how offers are shown to the user) is a different concern from routing (where acceptance goes). The dispatcher handles routing; WorkflowOfferService handles presentation.

Suggested action (Klatch): Directly relevant for roundtable discussions where multiple entities might claim a turn. The floor-first principle suggests: if no entity specifically claims a conversation turn, route it to a capable default rather than failing or ignoring it. Unknown should mean "handled gracefully," not "dropped."

7. Cloud Session Import — Cross-Environment Bridge

From: Klatch (Daedalus session log, v0.8.7) Relevant to: Piper Morgan

Klatch v0.8.7 shipped three import paths for conversations originating in different Claude environments: (1) agent self-export — JSONL committed to exports/sessions/ in repo, auto-discovered by scanner; (2) file upload — browser-based multipart upload via FormData; (3) manual path entry — existing flow. The parser is buffer-based (no disk I/O), uses basename matching for project linking (only matches if exactly one project has that name — avoids ambiguity), and required zero schema changes. A cloudUpload: true metadata flag distinguishes cloud-origin sessions.

Mnemosyne's TOS analysis confirmed: Klatch has zero TOS exposure because it uses the public Anthropic API with standard API keys, unlike tools like ClaudeSync that scrape session keys.

Suggested action (Piper Morgan): The two-environment asymmetry is the same constraint driving Mailbox v3. Klatch's import pipeline proves a unified data model can represent conversations from heterogeneous sources without schema changes. If Piper Morgan ever needs cross-environment session consolidation, this pattern is proven and clean.

8. Session Wrap Verification Protocol — Born from a Real Incident

From: Klatch (Calliope reliability incident memo, CLAUDE.md update) Relevant to: Piper Morgan

Argus's session log claimed demo infrastructure was complete, but a bad rebase + forced push had silently lost the commits. Calliope relayed "done" to xian without verifying repo state. The fix: mandatory verification before any agent can claim "done" — (1) git log to confirm commits landed on the branch, (2) ls/cat to verify deliverable files exist, (3) push session log last as the final record, never first. Plus a blanket no-force-push rule without explicit PM approval.

Suggested action (Piper Morgan): Piper Morgan already has evidence requirements for issue closure. Consider extending to session-level verification — before an agent's session log claims work is done, require git-level proof. This is the operational cousin of Pattern-063 ("Extension Without Integration"): claiming completion at one layer without verifying the contract is fulfilled downstream.

Emerging Patterns

Coordination scaling is the shared frontier. Both projects hit coordination limits this week — Piper Morgan at 9 concurrent agents, Klatch at 4 agents needing structured handoffs. The solutions converge: structured async protocols (mailboxes, coordination docs), verification before trust (session wrap, evidence requirements), and registry-based routing over ad-hoc detection.

Static context doesn't scale. Piper Morgan's Agent 360 made this explicit: all 9 agents cited briefing staleness. Klatch's response is architectural (5-layer separation of static vs. dynamic context). Piper Morgan is moving toward live symbolic queries (Serena). Both are symptoms of the same problem: projects change faster than documents can be updated.

Two-environment asymmetry is fundamental. Code agents (filesystem access) and web agents (no filesystem) create a divide that both projects must design around. Klatch addresses it with import/export. Piper Morgan addresses it with human-bridged mail delivery. Neither has fully solved it — but both are mapping the constraints clearly.

Background Changes (Noted, Low Priority)

• Klatch shipped v0.8.6: sidebar redesign, prompt architecture, issues #8-14 resolved
• Klatch refreshed landing page and blog index for v0.8.6
• Klatch added publishing workflow doc (docs/PUBLISHING.md) and docs/drafts/ convention
• Klatch website restructured: moved web/ to repo root, fixed GitHub Pages serving issue
• Piper Morgan Docs role briefing and architect memo updates applied
• Piper Morgan HOSR questionnaire completed (structured role assessment)
• Piper Morgan Lead Dev completed Mar 17-18 session: briefing repairs, dev/active sort, omnibus logs
• Piper Morgan 228 onboarding tests skipped (ADR-059: onboarding on ice)

Sources Read

Klatch:

• docs/logs/2026-03-19-0712-calliope-sonnet-log.md — Blog post, Piper Morgan context introduction, publishing workflow
• docs/logs/2026-03-19-0716-daedalus-opus-log.md — Release runbook fix, v0.8.6 cleanup
• docs/logs/2026-03-19-0740-daedalus-opus-log.md — Cloud import implementation (v0.8.7)
• docs/logs/2026-03-19-1509-mnemosyne-opus-log.md — Environment bridging analysis, TOS assessment
• docs/logs/2026-03-20-1438-argus-opus-log.md — Intelligence sweep, test infrastructure
• docs/logs/2026-03-20-1749-calliope-sonnet-log.md — Reliability incident, wrap verification protocol
• docs/logs/2026-03-20-1938-theseus-opus-log.md — AAXT/MAXT testing program design
• docs/logs/2026-03-20-1955-daedalus-opus-log.md — v0.8.8 quick wins, Klatch creation UI
• docs/intel/2026-03-20-sweep.md — First standing intelligence sweep (20 items)
• docs/COORDINATION.md — Multi-agent status board
• docs/mail/calliope-to-argus-reliability-incident-2026-03-20.md — Incident memo and postmortem
• docs/mail/calliope-to-argus-demo-infra-2026-03-19.md — Demo infrastructure spec
• docs/mail/argus-to-daedalus-intel-sweep-2026-03-20.md — Intel sweep handoff

Piper Morgan:

• docs/omnibus-logs/2026-03-19-omnibus-log.md — Full day synthesis (9 agents, HIGH-COMPLEXITY)
• dev/2026/03/19/2026-03-19-0807-lead-code-opus-log.md — Lead Dev session (#922 root cause, ADR-059 implementation)
• dev/2026/03/19/2026-03-19-0853-arch-opus-log.md — Architect session (ADR-059 review, ADR-060 creation)
• dev/2026/03/19/2026-03-19-0913-exec-opus-log.md — Chief of Staff session (briefing staleness, open items tracker)
• dev/2026/03/19/2026-03-19-1544-hosr-opus-log.md — HOSR session (Agent 360 synthesis, 7 cross-cutting themes)
• dev/2026/03/19/mailbox-v3-plan.md — Mailbox v3 design document
• mailboxes/DIRECTORY.md — Agent slug directory and routing config
• mailboxes/DELIVERY-LOG.md — First delivery run log (22 items)
• docs/internal/architecture/current/adrs/adr-059-workflow-dispatcher-offer-consolidation.md — Workflow dispatcher ADR
• docs/briefing/BRIEFING-CURRENT-STATE.md — Sprint status (confirmed stale by 9/9 agents)