The novel thing
What makes this AGI-ish is not a giant monolithic agent. It is a cognitive supply chain that turns reality into typed context, routes it through specialised judgement, and compounds from every run.
1. Omnipresent perception
Crons continuously sync the digital environment into source-near stores: Gmail, Calendar, Open Tabs, audio memory, cron outputs, GBrain, Signal Radar, finance, messaging, collaborator feedback.
2. Typed cognitive material
Extractors convert records into assets such as evidence, attention priorities, deadlines, relationships, uncertainty, boundaries, open loops, proof, feedback, outcomes, and source health.
3. Faculty-shaped judgement
Multiple specialist agents read the same substrate through different cognitive lenses, then a personal orchestrator suppresses duplicates, applies policy, and surfaces only what should matter.
The full category stack
Each stage answers a different kind of question. If these questions improve, the whole system becomes smarter when the underlying LLMs become smarter.
Reality surfaces
Gmail, Calendar, Telegram/SMS, Open Tabs, audio, GBrain, market/social feeds, cron outputs, finance, code, web.
System state
Services, crons, health checks, logs, registries, source freshness, model/tool availability.
Adapters
Read-only first connectors that know auth, privacy, idempotency, provenance, and safe mutation boundaries.
Heartbeat crons
Deterministic polling into raw/source-near databases. Code owns collection; agents own significance.
Source-near stores
Integration heartbeat SQLite, GBrain/Postgres, Open Tabs SQLite, audio SQLite, raw logs, run artifacts.
Context models
Asset builders convert source records into reusable cognitive objects with ids, source refs, freshness, caveats.
Judgement lenses
Focus, memory, pattern synthesis, execution, self-healing, question-asking, integration discovery, and more.
Collaborator
Merge, rank, suppress, route, ask, act locally, propose, delegate, and learn from feedback.
Context assets are the working memory objects
They are not summaries for their own sake. They are structured affordances for agent cognition: small, typed, provenance-bearing packets that can be retrieved by faculty-specific skills.
Asset types currently emitted
evidence 160uncertainty 127affordance 109attention priority 101boundary 87relationship 70temporal state 42deadline 34open loop 32feedback / proof / outcome 51waiting on 12source health 9Sources currently represented
Gmail 340Open Tabs 156GBrain 100cron outputs 56collaborator feedback 51calendar 50audio memory 48integration heartbeat 18signal 16finance 4messaging 4The faculty architecture
Faculties are cognitive organs. Their value is the tension they manage, the substrate they improve, and the verdict contract they obey: surface, suppress, question, repair, expand, execute.
Discernment
Substrate: Signal Radar filters, feedback ledger, suppression memory.
Focus
Substrate: Open Tabs, temporal state, deadlines, prior-run context.
Knowledge
Substrate: GBrain pages, facts, links, timeline, hot memory.
Pattern recognition
Substrate: context asset graph, GBrain, run artifacts, contradiction probes.
Creativity
Substrate: docs publishing, templates, prior successful artifacts.
Competence
Substrate: Hermes skill library, goal contracts, verification commands.
Conscientiousness
Substrate: policy gates, autonomy lanes, cron cadence, Agent Manager.
Self-improvement
Substrate: collaborator feedback, eval files, run manifests, capability scorecards.
Latest live faculty roster
Agent Manager, Cross-Scope Value, Execution, Focus/Open Tabs, Institutional Memory, Integration Discovery, Knowledge Gaps, Memory/GBrain, Pattern Synthesis, Procedural Memory, Question Asking, Self-Expanding, Self-Healing, Self-Upgrading, Signal Radar. The latest observable run records each as faculty-hermes-real-agent with output artifacts and contract validation.
The orchestrator is the prefrontal cortex
The Personal Orchestrator does not just call agents. It constructs an evidence packet, retrieves context assets, runs faculties, validates contracts, synthesizes, suppresses repeats, applies autonomy lanes, and emits an attention-protected collaborator surface.
Inputs
Raw/source-near heartbeats, GBrain, context assets, prior runs, feedback, system health.
Faculty judgement
Specialists produce verdicts, self-assessments, done conditions, source refs, and risk classifications.
Policy gates
Read-only, local artifact, local write, approval required, external side-effect blocked.
Personal Orchestrator
one collaborator voice from many cognitive lenses
Outputs
Telegram nudge, local artifact, goal packet, approval card, no-op, or source repair plan.
Memory writes
Feedback, facts, skills, GBrain pages, suppressions, proof artifacts, source coverage.
Proof
Manifest, AGENT_RUNS, INTELLIGENCE_REVIEW, CONTEXT_ASSETS, synthesis JSON, collaborator output.
Why this can feel intelligent
The user does not experience “15 agents.” The user experiences one collaborator that noticed something, remembered what happened last time, suppressed duplicates, gave a grounded next move, and knew when not to act. That is an intelligence simulation through harness design.
How it self-improves over time
The self-improvement loop is already visible, but it should become more explicit and more evaluative.
Feedback becomes policy
Connor replies useful / noisy / wrong / do_this. The system records card ids, suppresses repeats, and changes future ranking.
Runs become training data
Every run emits evidence, prompts, outputs, synthesis, and an intelligence review. This is an eval corpus for future faculty prompt and routing changes.
Skills become procedural memory
Repeated successful actions can be converted into Hermes skills: durable competence that future models can invoke.
GBrain compounds meaning
Exact user phrasing, facts, timelines, links, contradictions, salience, and trajectories become long-term semantic substrate.
Context assets compound perception
New integrations do not just add data. They add new asset classes, retrieval skills, and faculty affordances.
Crons make intelligence omnipresent
The system thinks in ticks: 5m reactive loop, 30m context extraction, 4h orchestrator, daily/weekly value loops, maintenance and health.
Gaps and simplification candidates
The architecture is strong. The next frontier is making the learning loops less implicit and the reality model richer without increasing nudge volume.
1. Faculty self-suppression is still too weak
The latest review scored self-suppression 3/5: 10 of 15 faculties emitted surface/question/action before synthesis. Synthesis protects Connor, but faculties themselves should learn stricter “nothing new” criteria.
2. Context asset semantics are still shallow
Many assets are typed from source records, but the asset ontology should mature: confidence, reversibility, owner, temporal horizon, decision impact, dependency, entity grounding, and expected half-life.
3. Outcome closure is underdeveloped
The system can propose, suppress, and log feedback. It needs stronger tracking from proposal → action → proof → outcome → memory update → policy change.
4. Source coverage is uneven
Gmail and Open Tabs are rich; finance and messaging are currently thin. Browser/SaaS activity, code activity, docs usage, location/context, sleep/health, and relationship-state signals are mostly absent.
5. Model-aware routing is not explicit enough
Simple extraction can use cheap models or deterministic code. Cross-source synthesis, counterfactual planning, and sensitive collaboration need stronger reasoning models and confidence-aware escalation.
6. The system needs a cognitive debugger
When a nudge is bad, Connor should be able to inspect: which source, which asset, which faculty, which prompt, which suppression rule, which prior feedback, which final ranking decision.
Where stronger reasoning models help most
This is the key harness-engineering test: if models get smarter, the system should get smarter because the hard questions are already isolated into the right stages.
Reality context that would make it more intelligent
Do not add sources because they are available. Add sources when they improve perception, attention, memory, action, or learning.
Digital work state
- GitHub/Linear issues, PRs, CI failures
- IDE/project activity summaries
- Docs usage and artifact feedback
- Browser/SaaS sessions as high-level activity, not raw surveillance
Communication reality
- Richer Telegram/SMS/WhatsApp exports
- Relationship follow-up state
- Commitments and waiting-on across channels
- Outbound draft approval workflow
Personal physiology/context
- Sleep, exercise, energy, calendar load
- Travel/location/timezone context
- Stress and focus budgets
- Capture friction: voice, quick notes, one-tap feedback
Financial/admin reality
- Bank exports and bills as read-only evidence
- Deadlines, subscriptions, cash timing
- Approval-gated actions only
- Proof of completed admin loops
External world model
- X/HN/RSS/podcast/YouTube market radar
- Exact-link reconciliation
- Competitor/company/entity trajectories
- Source reliability and novelty scoring
Internal wisdom model
- Connor’s frameworks and exact phrasing
- Contradiction/drift detection
- Preferences and decision patterns
- Repeated lessons turned into skills/policies
Next design moves
Small slices that increase intelligence without building another inbox.
Formalise a context asset ontology v2
Add confidence, owner, time horizon, reversibility, source reliability, decision impact, dependency links, and expected expiry.
Add faculty eval packs
Every faculty gets golden examples: should surface, should suppress, should ask, should route to local artifact. Run these after prompt/source changes.
Build proposal lifecycle tracking
Track proposed → approved/deferred/ignored → delegated/executed → proof → outcome → memory/policy update.
Create a cognitive trace viewer
A docs-style page per orchestrator run: source records → assets → faculty packets → verdicts → synthesis → final output → feedback.
Introduce model-tier routing
Cheap deterministic/LLM extraction for simple assets; stronger reasoning for cross-source synthesis, sensitive decisions, and self-improvement diffs.
Strengthen reality coverage carefully
Prioritise finance, messaging, code/work state, and browser/SaaS activity with privacy-preserving, read-only, source-near storage first.
Bottom line
The architecture is novel because it treats intelligence as a compounding socio-technical loop: perception from reality, memory with provenance, specialised judgement, policy-gated action, and feedback-driven improvement. It does not need to claim “true AGI” to be valuable; it successfully simulates many collaborator properties that people associate with general intelligence: remembering, noticing, prioritising, connecting, acting, learning, and being present over time.
Evidence base reviewed
Architecture-level facts only; no secrets or raw private content included.
- Context asset index summary: 843 assets across 11 sources.
- Source coverage report: all declared sources have source-near records, extractors, assets, and interfaces.
- Source adapter registry: read-only-first integration contracts and mutation boundaries.
- Latest Personal Orchestrator run manifest and AGENT_RUNS.
- Latest INTELLIGENCE_REVIEW: capability scores and self-suppression gap.
- Latest COLLABORATOR_OUTPUT: one final card, many suppressions, feedback contract.
- GBrain stats/health: pages, chunks, links, timeline, embeddings, brain score.
- Active cron list: sync, extraction, maintenance, watchdogs, digests, orchestrator runs.
- Existing docs.cforsyth.com artifacts: Making Agents Intelligent, Personal Intelligence Primitives, Personal Orchestrator MVP Architecture, System Registry, and related architecture docs.