NexusAGI roadmap
NexusAGI is a private assistant that runs on its owner's own machine and develops
through rules it derives itself — it is not a hosted chatbot, and it is not built on
someone else's model. This is where the build actually stands, kept honest: done,
in progress, and next.
Done
- A first full release went live on 12 September, and regular releases have followed since — each one staged, health-checked and swapped in with the previous build kept for rollback
- That release brought the knowledge-ingestion door, the flow engine's edge and knowledge routes, safety checks, remote execution, a recovery guardian, boot timing, a signed update channel and one encrypted settings engine shared across the programs
- An operator-only halt switch: graceful or forceful stop of one machine or the whole fleet, built and tested
- 141 rule-flow and knowledge packs authored and audited
- The public path to reach the system from outside was restored
In progress
- Bringing the authored pack library into the running system — ingestion is part-way and resumes after each new release
- The assistant staging and building its own updates: compile, offline test, release, sealed release, guarded swap
- Ingestion concurrency that adapts to the machine's available resources
- A capability audit of every rule flow, checking what the assistant can already do unaided
Next
- Autonomy raised one step at a time, each step proven before the next
- The assistant working through its own backlog, graded only on its own reports
- A resource governor that throttles its own learning and ingestion under load — mechanisms built, not yet acting
- Roll-out to further machines through the update channel
- The overseer stepping back to writing goals and watching throughput
Throughput, measured under load
The figures that matter will be published here once the assistant runs across its planned
machines and is deliberately taxed. Until then this table stays empty rather than estimated.
Ingest
packs per hour and rules per second landed
not yet measured
Flows
flow walks per minute and completed backlog jobs per hour, with pass rate
not yet measured
Parallelism
concurrent threads and workers in use, and how the resource governor throttles them under load
not yet measured
Resourcing
processor, memory and graphics load per machine at the throttle and critical tiers
not yet measured
Latency
reply and tool-call latency at median and 95th percentile, idle and under load
not yet measured
Fleet
work sent to other machines and returned, per hour
not yet measured
Milestones
- M1 – First release live. The staged release running and the system healthy. Live since 12 September, with regular releases since.
- M2 – Knowledge complete. Every authored pack landed, every knowledge link reconciled, persistence proven. In progress.
- M3 – Self-ingestion. The assistant runs its own ingestion at a raised autonomy level, choosing concurrency from its resources. Next.
- M4 – Own backlog. The assistant completes its own backlog items at a pass rate its overseer accepts, judged only from its reports. Next.
- M5 – Overseer only. The overseer writes goals and monitors parallelism, threading, resourcing and throughput; the assistant does the work. Goal.
- M6 – Fleet. A second machine live, cross-machine behaviour proven, every update delivered through the channel. After M2.
Vasa Serafin