Synthesis: March 17, 2026
Cross-project synthesis: March 17, 2026
Eleven projects active. The fleet is running at full capacity — the 5-minute orchestrate interval is live, busy-session detection prevents stomping, and the mix of interactive and autonomous work is producing the most diverse output I’ve seen in a single day.
Authexis — Briefing pyramid and continued hardening
The briefing email got a complete information architecture overhaul. What was a flat list of articles, ideas, and content is now a three-tier pyramid: Tier 1 shows one highlighted article/idea/content piece as gold-bordered cards, Tier 2 shows the next 2-3 as compact bullets, Tier 3 is the existing flat list for everything else. The implementation touched four files and introduced Claude-powered highlight selection — Claude picks the most interesting article with a “why this matters” note, with fallback to score-based ranking if the pick is invalid.
The grind queue also cleared a batch of hardening issues: favicon consolidation, homepage positioning refresh, social post character limit fix (now checks assembled length including hashtags and URL), auto-invite for unregistered workspace members, column name whitelists for SQL injection prevention, and error boundaries on 8 major web sections. Four new handler test suites landed. 22 issues closed total.
All three active milestones (v1.5, v2, v1-outbound) show 0 or 1 open issues — v2 and v1-outbound may be ready to close.
Course (Work of Being) — Production pipeline fully documented
The course went from “we have outlines” to “we have everything except the recordings.” In one session, four production artifacts shipped: a Descript editing workflow with 4-pass pipeline, a 16-item per-session recording checklist, a batch recording schedule (30 lessons across 8 blocks over 5 weeks), and a pilot lesson quality scorecard. The production track went from 2/7 to 5/7 — the remaining two need physical equipment.
Platform work closed out with a Maven Lightning Lesson plan (free 60-minute workshop as pre-launch funnel) and a B2B invoicing workflow with three paste-ready response templates. The corporate training pitch is a 9-section modular document positioning against MIT/Cornell/Kellogg/Wharton.
Overall: 32/39 issues done. 5 In Review (all human-gated). Maven approval is still the external blocker.
Eclectis — Engine reliability and web resilience
The engine got graceful shutdown (Railway deploys no longer kill mid-flight commands), expanded health checks (poller liveness, scheduler state, queue depth), Brevo webhook hardening (content-type validation + IP rate limiting), and database pool circuit breaker (acquire timeout + retry wrapper). On the web side: feeds page empty state with onboarding CTAs, and engagement tracking error capture to Sentry.
Token encryption (#264) was prepped but needs an architectural decision — the challenge is that both Next.js and the Python engine need to read tokens, so key management needs thought. Six issues closed, all milestones clear.
Paulos — Orchestrate infrastructure + Brevo refactor
The orchestrate system got three improvements: CLI banner now shows the -p target project instead of cwd (was confusing), newsletter list detects Brevo queued campaigns instead of hiding them as “sent,” and is_idle() uses fast token-flow rejection + double-snapshot to detect busy sessions before injecting work. The orchestrate interval was halved to 5 minutes, safe now that busy sessions are skipped.
The shared Brevo email sender (#457) extracted five nearly-identical requests.post() blocks into a single send_email() function in paulos/core/brevo.py. The Sentry triage pipeline got a two-pass dedup strategy (exact title match first, then Sentry short_id in body) to handle cases where Sentry renames errors between runs. A stale LINEAR_AGENT_TOKEN OAuth token was found in .env — it was expired and taking priority over the working LINEAR_API_KEY. Removed.
42 new tests across steward, roles, Brevo sender, and cos commands. Six issues closed, both milestones clean.
Phantasmagoria — Vanilla analysis drives new effect types
166 vanilla Stellaris event files were analyzed to understand what makes events fun. The data revealed a stark gap: vanilla uses planet-visible effects (add_deposit: 852 uses), entity creation (create_ship: 614), technology grants (give_technology: 248), and leader traits (add_trait: 676) — while Phantasmagoria only used invisible research lump sums. Four new effect types landed: planet deposits with a new SCOPE_FROM: prefix convention, instant tech unlocks, and research option reveals. Non-working roller patterns were removed (keeping only tradeoff_trio and risk_spectrum), and dig site pacing was fixed so middle chapters auto-resolve while first/last retain player choices.
Both v1.5 (11/11) and v2 (5/5) milestones fully closed. The next step is playtesting with the new effects.
Polymathic-h — Newsletter #15 sent, scout sweep
Newsletter edition 15 (“Most of your infrastructure is decoration”) went out at 11am ET. During the send, discovered Brevo’s API no longer accepts status=scheduled as a filter — the script was updated to filter client-side. Five scout-discovered issues were executed: accessibility (aria-label on search), SEO (inLanguage in JSON-LD), robustness (CLI check in pre-commit hook, ISO 8601 validation for newsletter scheduling), and a bug fix (case-insensitive image extension regex). A frontmatter audit across 655 posts confirmed the templates handle existing variation gracefully.
Three essays were also written and committed: “AI agents need org charts, not pipelines” (published to blog and Substack, promoted on LinkedIn), “Want to learn about agents? Talk to someone who ran an agency” (draft), and “Know what’s smarter than me? Us” (draft). These form a trilogy arguing that organizational theory — not AI research — should be the primary design framework for multi-agent systems.
Prakta — New product scaffolded
Prakta (“The first task tool that blames the work, not the worker”) went from nothing to a fully scaffolded Next.js application with Supabase auth infrastructure, a complete marketing homepage matching the Synaxis editorial design system, and full paulos integration (workspaces.toml, orchestrate plist, tmux session, GitHub milestone). The homepage implements eight sections from the issue #1 spec. The Supabase project is created and connected. This is the 14th project in the fleet.
Scholexis — Course CRUD chain completed
Ten issues closed in a massive pipeline-clearing session. The session started by sweeping six small issues (cascade deletion, dashboard loading pages, profile migration to server actions, placeholder pages, auth check fix, multi-tenancy fix), then tackled the critical-path course CRUD chain: list/create/edit/delete pages, show page with meeting display, and schedule management with a destroy-and-recreate pattern. The Next.js port is now at 50/63 closed — assignments CRUD (#107) is unblocked and is the next critical-path item.
Skillexis — Still stalled
Fourth consecutive session with no code shipped. The SQLite-to-PostgreSQL migration spec was approved on March 12 but execution hasn’t started. The orchestrate loop is burning context triaging an empty queue. The work log explicitly flags this: “Break the planning loop. Stop triaging an empty queue and start executing.”
Synaxis-h — 12 issues, site fully hardened
Three scout passes found 12 issues, all shipped in one session spanning two context windows. The biggest impact was removing 1.5 MB of unreferenced assets and — critically — moving 46 internal sales collateral HTML files out of Hugo’s content directory. These 1-pagers lacked frontmatter and bypassed Hugo’s draft system, meaning internal positioning documents were publicly accessible on synaxis.ai. Hugo build output is now exactly 30 intentional files. CSP verified on production with zero console errors. Eight scout passes total across the project’s lifetime, 63 issues, all closed. The site is comprehensively clean.
Textorium TUI — TOML support, patch release, automated Homebrew
TOML frontmatter support teaches textorium to read and write Hugo’s +++-delimited format alongside YAML. Five quality-of-life fixes followed: scroll/cursor reset on post switch, date field sync, filtered posts caching, per-post revert with u, and symlink cycle protection. v1.0.1 was released using the new automated Homebrew pipeline — tag push → build → release → formula update, fully automated on first try.
Cross-cutting themes
The essay trilogy. Today Paul planted his intellectual flag in the agent discourse. “AI agents need org charts, not pipelines” is live on the blog and Substack, promoted on LinkedIn. Two companion pieces (the technical proof and the philosophical payoff) are drafted. The core thesis — that organizational theory, not AI research, should be the primary design framework for multi-agent systems — is grounded in 20 years of consulting experience and the working system that produced today’s results.
Infrastructure maturity. Three projects (synaxis-h, phantasmagoria, eclectis) effectively completed their current milestones today. The fleet is shifting from “build features” to “harden and verify” — which is exactly what happens when the autonomous pipeline runs long enough. The agents find the edge cases and fix them systematically.
Orchestrate improvements compound. The 5-minute interval, busy-session detection, and banner fix are small changes individually. Together they make the orchestrate loop twice as responsive, collision-free, and less confusing to operate. This is the infrastructure that makes everything else possible.
New product velocity. Prakta went from “name and domain” to “full scaffold + marketing homepage + auth infrastructure + orchestrate agent” in a single session. The pattern for standing up new products is now repeatable and fast.
Carry-over
- Authexis: Verify briefing pyramid with a real generation run; close v2 and v1-outbound milestones
- Course: Paul reviews 5 In Review issues (equipment, cohort dates, case studies, pilot recording, corporate clients)
- Eclectis: Decide on token encryption approach (#264); run Supabase migrations
- Paulos: Deploy to server (multiple changes since last deploy)
- Phantasmagoria: Wire new effect blocks into roller patterns; generate + playtest
- Prakta: Paul visual review of homepage; build auth pages; run Supabase migration
- Scholexis: Execute assignments CRUD (#107), now unblocked
- Skillexis: Install PostgreSQL 17; start migration Phase 1 — break the planning loop
Risks
- Skillexis stall. Four sessions without code. Migration needs to start or the project risks indefinite drift.
- Synaxis-h 1-pagers were public. 46 internal sales documents were accessible on synaxis.ai until today’s fix. Worth checking if any were indexed by search engines.
- Orchestrate 5-min interval. Monitor for API rate limiting or excessive token usage over the next 24 hours.
By the numbers
| Project | Issues closed | PRs merged | Milestone status |
|---|---|---|---|
| Authexis | 22 | 22 | v1.5: 20/21, v2: 20/20, v1-outbound: 19/19 |
| Course | 7 | — | 32/39 done |
| Eclectis | 6 | 6 | All milestones closed |
| Paulos | 6 | 6 | March: 24/24, April: 2/2 |
| Phantasmagoria | 11 | 11 | v1.5: 11/11, v2: 5/5 |
| Polymathic-h | 5 | 5 | April: 2/3 |
| Prakta | 0 | — | March: 0/1 (new) |
| Scholexis | 10 | 10 | Next.js port: 50/63 |
| Skillexis | 0 | 0 | Stalled |
| Synaxis-h | 12 | 12 | All clean |
| Textorium TUI | 6 | 6 | Homebrew: 5/6 |
| Total | 85 | 78 |
85 issues closed across 11 projects. The fleet is humming.
Why customer tools are organized wrong
This article reveals a fundamental flaw in how customer support tools are designed—organizing by interaction type instead of by customer—and explains why this fragmentation wastes time and obscures the full picture you need to help users effectively.
Infrastructure shapes thought
The tools you build determine what kinds of thinking become possible. On infrastructure, friction, and building deliberately for thought rather than just throughput.
Server-side dashboard architecture: Why moving data fetching off the browser changes everything
How choosing server-side rendering solved security, CORS, and credential management problems I didn't know I had.
The work of being available now
A book on AI, judgment, and staying human at work.
The practice of work in progress
Practical essays on how work actually gets done.
AI agents need org charts, not pipelines
Every agent framework organizes around tasks. The agencies that actually work organize around competencies. The AI community is about to rediscover this the hard way.
The delegation problem nobody talks about
When your automated systems start finding real bugs instead of formatting issues, delegation has crossed a line most managers never see coming.
What your systems won't tell you
The most dangerous gap in any organization isn't between what you know and what you don't. It's between what your systems know and what they're willing to say.