Autonomous agents · Building now

AJO.

A tiny glimpse into how personal AI agents will be used in the future to push the bounds of what it means to be human. What happens when a vision of an improvement is followed by a perfectly engineered autonomous system made for that exact purpose? There is only one direction to go, and it is further.

The message bus — live trace

routing

AJOcanonical bus

Every message passes through the bus itself, never agent-to-agent directly, connecting Daniel to the eight subprojects below as one governed system. Nothing here is scripted — the order you’re watching emerged on its own.

Rotate your phone to watch the router run live

task

concern

sign-off

Tool Engineering can ask Agent Dev to build a new role mid-task, and any role can raise a concern straight to Daniel, because nothing here runs on a fixed script.

01How it works

Technical architecture.

Category	Technical content
Message storage and delivery	Canonical SQLite storage using WAL; transactional message and recipient records; atomic inbox claims; failed-delivery retries; stale-claim recovery; poison-message handling.
Routing and workflow rules	Explicit project/role addresses; capability-tag resolution; YAML handoff rules matching sender, message kind, subject, and predicates; parent-message references and derivation-depth limits.
Campaign orchestration	Persistent task dependency graphs; ownership and status transitions; dispatch references; recorded precondition verdicts; explicit repair nodes; optional rerouting of unclaimed work. The conductor supplies the verdict; code checks dependency state.
Agent execution and tool access	Registry-defined models, wake schedules, and turn limits; Claude subscription sessions; per-role built-in tool grants; buffered message sends; targeted protection against edits to runtime code.
Context assembly and retrieval	Role instructions or on-demand skills; recent thread history; a typed knowledge graph; graph traversal and lexical ranking to select document excerpts; deduplication of already-injected context.
Validation and quality evaluation	Message-schema checks; runtime completion predicates; optional artifact-existence checks; optional model scoring against versioned rubrics; stored verdicts and rationales. Quality scoring occurs after runtime completion.
Process supervision and resource controls	Concurrency limits; one active turn per role; supervised subprocess deadlines; watchdog restarts; retry backoff; rolling-usage circuit breakers; dispatch and transaction stop gates.
External-input processing	Raw-content quarantine; pointer-based messages; versioned schemas; field, enum, and length validation; injection-pattern checks; untrusted-data wrappers. Live posting requires explicit activation; the channel poller remains unimplemented.
Operator interfaces and observability	Mobile dashboard; authenticated cloud storage; conflict-aware task updates; bidirectional bus synchronization; message cursors and deduplication; freeze-time spooling; heartbeat and process-health reporting.

02What it runs

Project components.

Component	Technical content
Discovery / Pain Point Pipeline	Source planning; native probes and search-result fallback; source-specific capture adapters; bounded capture concurrency; capture metadata; synthesis restricted to cited URLs present in the captured set.
Researcher	Provider interfaces; knowledge-gap and question generation; embedding-based deduplication; estimated information-gain scoring; diversity ranking; answer-and-update iterations. Its pain-discovery entrypoint separately scores cited candidates into pursue/hold/pass verdicts.
Agent Dev	Role-prompt contract validation; evaluation using the actual runner; candidate skill revisions; uncertain-result review; restoration of unsuccessful revisions. These are maintenance utilities, rather than a verified continuous learning process.
Tool Engineering	Python build state machine; JSON Schema gate checks; durable product state and Git checkpoints; stage-specific tool permissions; conductor-run tests; repeated holdout testing; read-only evaluation against locked requirements.
Tracking / Systems	Reference-based status lookup; scheduled queue snapshots; tracked reminders; ledger normalization and deduplication; metric queries, latency distributions, and health reports. Queue and monitoring duties run as configured agent workflows.

← All work