Synthesis: April 3, 2026

Five projects shipped today, and for the first time the whole fleet moved like one organism. Paulos built the infrastructure. Synaxis AI used it. Diktura went from scaffold to real product. Eclectis shipped features from client requests. Authexis cut its own bills. The thing that made it possible wasn’t any single feature — it was the fact that every session could reach into every other session and ask questions.

Paulos — Infrastructure and observability

Nine issues closed. The headline is the API dispatch path — three issues (#671, #672, #673) that give the orchestrate loop a third execution method alongside tmux injection and claude -p. Set dispatch_method = "api" in a workspace config and the loop calls skopos.py (the API-backed agent runtime) instead of shelling out to the subscription-backed CLI. This is the ToS insurance policy: if Anthropic enforces the subscription-automation restriction, every project can flip to API dispatch without code changes.

Two new CLI commands: paulos mail check scans the Fastmail inbox via JMAP and categorizes emails (deploy alerts, errors, prospect replies, human), and paulos cos audit-docs audits context file sizes across the fleet and flags bloat. The audit tool was immediately dogfooded — paulos’s own PRODUCT.md and DECISIONS.md were 10.5K chars too heavy (Features reference duplicated the Layer 2 inventory; 4 stale decision entries). Cut to 48K total, down from 59K.

The biggest discovery: Sentry SDKs have been installed on authexis and eclectis since March, reporting to the synaxis-ai org. But no project slugs were wired into the triage pipeline. Thirteen unresolved errors sitting there, nobody looking. Fixed the wiring — both projects now scanned by paulos sentry triage. Infrastructure research across all four platforms (Sentry, Vercel, Railway, Supabase) confirmed webhook capabilities for deploy-failure alerting.

Model routing was documented: Heavy (Opus for judgment), Standard (Sonnet for well-specced work), Light (Haiku for classification). The implementation already existed — the policy just needed writing down.

Synaxis AI — First real engagement, complete architecture overhaul

The day’s biggest story. The agency ran its first real engagement end-to-end: Authexis was onboarded as a client, Sydney produced an onboarding document, Alex reviewed it, Paul approved with comments, Alex ran the learn cycle, Sydney polished, Trina delivered via tmux, the client (authexis session) approved with comments, the document was archived. Every status in the model was exercised. The methodology was updated eleven times during the engagement.

The architecture was overhauled from the ground up. Five agency skills now drive everything: /agency-onboard (Notion + Slack infrastructure), /agency-engage (scaffold engagement from playbook), /agency-loop (Trina reads the board, picks an action, dispatches), /agency-work (persona acts on a deliverable), /agency-ask (cross-session tmux questions). The old paulos orchestrate agency dispatch was retired — skills-only, running in Claude Code sessions.

The status model grew from 8 to 12 states. Review responses at each gate: accept, accept with comments, reject. Two post-review paths: Final learnings → Final tweaks → Done (for approvals), Revisions → New (for rejections). Voice gate was removed — Paul always reviews.

Two blog posts drafted and heavily edited: “I ran my AI agency’s first real engagement” (blow-by-blow) and “The agency in 2026: AI for execution, humans for judgment” (architecture through objection cascade). Both at draft: true on polymathic-h.

Diktura — From scaffold to real product

Ten issues closed. The entire roadmap feature was built: kanban list page (#90), detail page with inline editing (#91), public widget voting API (#92). That closes out the Rails port — Diktura now has feature parity across feedback, changelog, broadcasts, app users, roadmap, and voting, all in the new Next.js + Python stack.

A scout run found and fixed four issues in the same session: unbounded roadmap query (#93), 16 missing changelog tests (#94), responsive kanban grid (#95), and loading skeletons for 5 dashboard pages (#96). The landing page was replaced with a real SaaS marketing page (#98), the roadmap link was added to nav (#99), and the login page went from placeholder to working Supabase auth (#100).

The product vision expanded significantly. Paul clarified that Diktura isn’t “feedback + roadmap + changelog” — it’s a unified customer interaction platform. Support, KB, inbound email/SMS, newsletter, Sentry errors, PostHog analytics, all attached to one customer identity across multiple apps. PRODUCT.md was rewritten.

Infrastructure win: the Railway engine was fixed. The DATABASE_URL was missing, and the Supabase pooler host for Ohio projects is aws-1-us-east-2, not the aws-0-us-east-1 that docs suggest. Engine is now live and polling.

Eclectis — Feature delivery from client onboarding

Four issues closed, all driven by the Authexis client onboarding process. Manual briefing mode (#661) lets users toggle between auto and manual briefing, with checkboxes on article cards and a “generate briefing” button. The engine’s briefing_generate handler accepts optional article_ids to skip auto-scoring.

Three onboarding improvements from #660 decomposition: file upload (#663) for .txt/.docx files on onboarding step 1, and LinkedIn profile URL (#665) stored in preferences JSONB for manual Claude cowork review. Both avoid database migrations by using the existing JSONB column.

Slack bot posting was set up — the Eli bot token is in .env and tested. Read via MCP plugin (Paul’s account), write via bot token (Eli bot identity).

Authexis — Cost optimization and agency onboarding

The Vercel bill got cut by ~96%. An audit revealed Turbo build machines ($0.126/min) with auto-deploy on every push — roughly 22 production builds per day. Switched to Standard machines ($0.014/min) and set the ignored build step to exit 0 so deploys only happen via vercel --prod. Estimated monthly cost drops from ~$99 to ~$3.43 in build minutes. The /ship skill was updated to include vercel --prod so deploys still happen at ship points.

On the agency side, the product session answered two rounds of onboarding questions from Synaxis AI — Sydney’s product interview (8 questions about positioning, voice, pricing) and Trina’s doc review. Both answered from PRODUCT.md and DECISIONS.md, with 4 open business questions flagged for Paul.

Cross-cutting themes

The fleet is one organism now

Today was the first time multiple sessions routinely talked to each other during work. Synaxis AI’s /agency-ask typed questions into the authexis tmux session, got answers from its PRODUCT.md, and used them to produce a client onboarding document. The authexis session didn’t know or care that it was serving an agency workflow — it just answered product questions. This is the multi-agent collaboration pattern working at production scale.

Observability gaps closing

Three independent discoveries today: Sentry errors accumulating silently (paulos), Vercel costs running unmonitored (authexis), and Railway healthcheck failing with no alert (diktura). The common thread: infrastructure that reports to email, and nobody reads the email. The fix is a three-phase pipeline: dashboard webhooks (quick), CLI polling (medium), webhook→issue automation (later). The Sentry pipeline is now reconnected. The Vercel/Railway webhooks are identified but need manual dashboard configuration.

Product visions expanding

Both Diktura and the Synaxis agency had significant product pivots today. Diktura went from “feedback tool” to “unified customer interaction platform.” Synaxis went from “paulos orchestrate agency” to a skills-only architecture running real engagements. In both cases, the pivot came from Paul engaging directly with the product session and making scope decisions that the agents couldn’t make on their own.

Cost consciousness

Authexis cut its Vercel bill by ~96%. Paulos cut context injection by 18%. The model routing documentation codifies the principle: use Sonnet for well-specced work, save Opus for judgment. The ToS hedge (API dispatch) is partly a cost question too — API credits are metered, subscription is flat. The fleet is getting disciplined about what it spends.

Carry-over

• Blog posts — two Synaxis AI posts at draft: true, close to final. Publication order and timing TBD.
• Diktura Vercel deployment (#101) — feature-complete, auth works locally, never deployed. This is the last barrier to a live product.
• Vercel settings — eclectis, prakta, dinly still on Turbo machines with auto-deploy. Paul needs to change in dashboard.
• Sentry backlog — 13 old unresolved errors across authexis (11) and eclectis (2). Pipeline catches new ones; backlog needs eyeball.
• Railway/Vercel Discord webhooks — 5-minute dashboard configuration, identified but not done yet.
• Authexis product launch engagement (#66) — 9 deliverables, parallel tracks. The real test of the agency system.
• Eclectis onboarding (#67) — second client through updated flow. Validates methodology generalizes.
• Notion cleanup — stray view tabs, verify status properties, clean up voice_gate remnant.

Risks

• diktura.io not deployed — feature-complete product with no production deployment. 83 unit tests, zero integration tests against real data.
• No continuous health monitoring — Railway healthcheck is deploy-time only. diktura-engine could be down with no alert.
• Sentry errors accumulating — 13 unresolved since March. Pipeline connected now, but backlog unaddressed.
• Notion API limitations — tables don’t render, linked views can’t be created, view filters don’t apply. Affects agency deliverable quality.
• Multi-session agency handoff untested — the Authexis engagement ran in one context window. Multi-session via work logs + Notion state hasn’t been proven yet.

By the numbers