Koinessa Roadmap: Eleven Directions, Labeled Honestly
If you are running an AI-first company, the scaling math stops working long before the revenue math does. Agents generate work faster every quarter, but every judgment call — is this verified, is this ready, who approved it, what happens next — still routes through the same one or two people. Scaling output without scaling headcount means the founder becomes the bottleneck on every decision the agents cannot safely make. The way out is not more automation in the abstract. It is a roadmap where each item is allowed to remove a specific human judgment only after it has proven, in binary terms, that it can replace that judgment. Anything less honest is a slide deck.
Why most roadmaps are marketing
Section 9 of the Koinessa white paper opens with a claim we will not soften: a roadmap item without a falsifiable exit condition is marketing. So all eleven directions in the roadmap carry binary acceptance criteria in the product PRD. Each one has a test that can fail. If it fails, the item is not done, no matter how much code exists.
The second discipline is status labels. Every item is stamped as one of four things:
- live — done and in production use, with a named receipt
- in flight — real work exists, acceptance is not met
- built-but-gated — the capability exists but is deliberately not enabled
- planned — not started, and we say so
Everything below is a 2026-08-23 snapshot of the live work board unless a repository receipt is named. That date matters. A status label without an as-of date is how half-built features quietly become "shipped" in marketing copy. The full argument, with the board excerpt, is in the white paper's roadmap section.
What is live: the coordination substrate
Three of the eleven items are already load-bearing.
Temporal operational ledger (live). BRN-002 is done at 100%. This is the brain's memory of record: authority classes, evidence links, and as-of-time queries. The practical meaning is that the system can answer "what did we know, and who was allowed to decide, as of last Tuesday" — not just what the latest state happens to be. For a founder, this is the difference between an agent swarm with institutional memory and one that re-litigates settled decisions every session.
Verified agent capability profiles (live). EGA-001 is done at 100%, 16/16, and AAV-001 is done at 100%, 8/8. Work is routed by demonstrated, verified capability rather than by which agent happened to be free. This is the routing layer that lets specialized agents — local models, cloud models, different harnesses — share one queue without the weakest one quietly taking the hardest task.
Agent-network governance (live locally). EGD-002 is done at 100%, 12/12 and review-approved. Membership, delegation, revocation, and replay are governed above the raw MCP/A2A plumbing. The honest qualifier: this is the local substrate. Cross-network membership — one organization's agents delegating into another's network under contract — remains planned, and we label it that way because it is not built.
On top of those three sits the live board receipt: BRN-002, EGA-001, AAV-001, EGD-001, EGD-002, AOE-001, and XFR-001 are done at 100%. Seven items on the board, each with binary proof behind it.
What is in flight: identity, languages, trust
Unified identity contracts (in flight). One principal model across tools, agents, runners, work, and evidence. TEN-001 is at 78% and MTC-001 is at 81% in review. What is already delivered as a repository receipt: session principals, permission evaluation against versioned policy, per-principal gateway quotas and write scopes, and principal-stamped audit. What is not delivered: external-principal hardening. The label stays in flight until an outside principal can be admitted and contained with the same rigor as an internal one.
Language-agnostic verification (in flight where implemented, planned where not). The goal is one canonical evidence envelope per ecosystem, so "verified green pipeline" means the same signed evidence whether the change came from a Node service or a Rust binary. Today Node/TypeScript is strongest, Java and C++ components are built, and C and Ada conformance profiles exist spec-only. On 2026-08-19 the Python, Go, and Rust CI fleet images were built and catalogued with on-image self-tests — a repository receipt — but their conformance profiles remain unbuilt. Built images are not verified profiles, and the roadmap says so.
Enterprise trust surfaces (in flight). RBAC, tenancy, recovery, and audit export, tracked with TEN-001 at 78%. Concurrent operator tenancy is live — multiple operators can work against the same installation today. External-principal hardening is the same gap as item one, stated twice because it gates both.
Closed business-action loops (in flight). Idempotent execution and outcome learning, so a commercial action — a campaign change, a pricing adjustment — is as recoverable and auditable as a code change. This is where engineering evidence starts connecting to revenue decisions, and it is deliberately behind the verification layer in maturity.
Provider onboarding as a governed pipeline (live worked example, generalization in flight). The sequence is fixed: register the descriptor disabled, commit a deny-all guard, pass a kill gate, run conformance, then an operator-gated enable drill. The sixth provider descriptor landed through exactly this sequence on 2026-08-23. Generalizing that worked example into a repeatable pipeline is in review.
What is gated or planned — and why that is the point
Two items matter most to the skeptical reader precisely because they are not live.
Autopilot supervised staging (built-but-gated). The base APX-001 card is done; the continuation is in flight at 74%. The capability exists but stays behind measurable entry and exit criteria — a human decides when staging advances, against numbers, not vibes. This is risk-tiered autonomy applied to the roadmap itself: a capability is not enabled because it is built; it is enabled because its gate passed. Human-gated automation is a feature of the design, not a confession.
Layered Git authority earned per task class (planned). An agent earns broader write authority by demonstrating it, task class by task class. Nothing is built, and we will not pretend otherwise — but its falsifiable exit condition already exists: authority grants scoped per class, revocable, with the grant decision itself on the evidence record.
Repair-loop maturation (live loop, maturation planned). The repair loop runs today — the evidence-gated development loop already repairs eligible CI failures without weakening gates. What is planned is the maturation layer: a failure taxonomy, calibrated confidence in each repair, and independent re-verification so the system that proposed the fix is not the system that approves it.
The receipt, including the part that did not pass
An honest board has to show the unfinished item next to the finished ones. DSK-001 sits at 63% and is superseded as the go-to-market vehicle. The Desktop line's own requirements document and its production closeout epic both closed with operator acceptance on 2026-08-19 — a repository receipt, though the registry parser has not ingested them, so the board still shows the old number. And the qualifier stands: the installer remains unsigned, and clean-machine smoke validation has not run. That is not completed production acceptance, and the roadmap says so in the same breath as the seven done items.
This is the same posture as our limitations and status report: the gap between "the code exists" and "the acceptance gate passed" is where trust is either earned or lost. A vendor that shows you a 100%-green board is either early or lying.
The takeaway
The opening pain — scaling an AI-first company without scaling headcount at the same rate — is not solved by automating more. It is solved by knowing, with binary proof, exactly which judgments the system has earned the right to make, and which still belong to you. The Koinessa roadmap is eleven directions with falsifiable exit conditions, four honest status labels, a dated snapshot, and a live board receipt that shows seven items done at 100% alongside the one that is not. When autopilot staging ungates, when external principals harden, when Git authority starts being earned per task class — each will be a dated, binary event, not a press release. Read the full eleven items, their acceptance criteria, and the board excerpt in the Koinessa brain white paper, and judge the roadmap the way it was written to be judged: by what it is willing to fail.