Personal Orchestrator MVP

A clean Hermes-native collaborator layer: integrations heartbeat raw/source-near data into one local database; many faculties inspect the same substrate; Hermes orchestrates faculty subagents; feedback and run reviews tune the system over time.

MVP state: live local run verifiedNo Hermes Control PanelNo product JSON middlewareDeterminism limited to capture + safety

Does it work like you think?

Yes: many faculties can read the same source layer

The current MVP stores Open Tabs, GBrain, Signal Radar, and Gmail source-near records in state/integration_heartbeat.sqlite. Faculty agents receive a shared evidence packet and can be extended to query the heartbeat DB directly.

Yes: Hermes is the orchestrator

runners/observability_run.py collects evidence, runs deterministic local checks, invokes real Hermes one-shot faculty agents, and writes auditable artifacts per run.

Partly: self-improvement exists as a loop, not yet intelligence

Feedback ledgers, suppressions, faculty eval files, and INTELLIGENCE_REVIEW.md now create the tuning surface. The next jump is converting Connor feedback into prompt/source/eval changes automatically but safely.

The shape

The important design change: raw integration heartbeat is the only deterministic lane. Everything above it is judgement, synthesis, risk policy, or feedback.

deterministic

Integration heartbeats

  • Open Tabs
  • GBrain
  • Signal Radar
  • Gmail readonly
  • future: Calendar, audio, finance, messaging
→
deterministic

Raw/source-near DB

  • poll + dedupe
  • provenance
  • privacy tier
  • retention policy
  • idempotency key
→
judgement

Faculty subagents

  • read the same evidence
  • apply different lenses
  • surface, suppress, ask, repair, expand
synthesis

Cross-faculty comparison

  • agreement/conflict
  • ranked collaborator moves
  • one useful question when blocked
→
policy

Action gradient

  • low-risk auto-fix
  • approval-required cards
  • forbidden actions
  • silence when not useful
→
learning

Feedback + reviews

  • Approve / Ignore / Edit / Defer
  • faculty evals
  • INTELLIGENCE_REVIEW
  • prompt/source tuning

Faculty subagents

Current roster in agents/faculties/<faculty>/. Each has a prompt, data contract, source policy, and eval file.

Integration Discovery

Find repeated source needs and propose new heartbeat integrations.

gbrainbacklogsession

Signal Radar

Separate useful external context from noisy feeds.

signal-radarx/youtube/spotify

Focus / Open Tabs

Protect attention, decisions, deadlines, and cognitive load.

open-tabscalendarbacklog

Memory / GBrain

Decide what to recall, save, correct, forget, or keep as source evidence.

gbrainaudio-memorytelegram

Pattern Synthesis

Find cross-source structures that no single source proves alone.

all evidencegbrainopen-tabs

Agent Manager

Convert opportunities into safe approval cards and suppression feedback.

backlogopen-tabsreactive events

Procedural Memory

Notice repeated workflows that should become Hermes skills.

run artifactsgit diffssession outcomes

Execution

Route approved work to local tools or delegated agents and verify done state.

approvalsbacklog

Self-Healing

Find broken collectors, crons, brittle loops, and safe repairs.

cronsource errorshealth

Self-Upgrading

Turn repeated failures and gaps into implementation slices.

capability scorecardfeedback

Knowledge Gaps

Name missing context when it changes the recommendation.

faculty outputsevidence packet

Question Asking / Epistemic Gap

Ask the single highest-leverage question before acting.

source-near recordspolicy conflicts

Self-Expanding

Propose safe new source/faculty scope when evidence supports it.

integration discoverycapability scorecard

Cross-Scope Value

Measure whether connecting sources actually produced value.

all facultiesoutcome metrics

Data / interface access

The interface is a source policy + heartbeat namespace, not a pre-filtered proposal API.

Source
Heartbeat data captured
Who reads it
Open Tabs
Items, decisions, deadlines, backlog-ish attention loops via /root/open-tabs/open_tabs.py list --json.
Focus, Agent Manager, Pattern Synthesis, Execution, Self-Healing.
GBrain
Recent salience, memory/concept/fact context, contradictions/gaps via GBrain CLI/MCP.
Memory/GBrain, Knowledge Gaps, Pattern Synthesis, Question Asking, Cross-Scope Value.
Signal Radar
Latest external signal artifact from market/social/media awareness state.
Signal Radar, Pattern Synthesis, Integration Discovery, Cross-Scope Value.
Gmail
Readonly recent metadata/snippets; full bodies should be fetched only when a faculty/run has explicit need.
Focus, Question Asking, Pattern Synthesis, Agent Manager, Self-Healing, Execution. Sends/archive/labels remain approval-gated.
Future: Calendar
Meeting load, deadlines, prep needs, conflicts.
Focus, Execution, Pattern Synthesis, Question Asking.
Future: Audio Memory
Raw thoughts, commitments, emotional salience, memory candidates.
Memory/GBrain, Pattern Synthesis, Question Asking, Agent Manager.
Future: Finance
Read-only cash/runway, bills, unusual spend, portfolio thresholds.
Question Asking, Pattern Synthesis, Conscientiousness-sensitive Agent Manager. Mutations forbidden without explicit one-off approval.
Future: Messaging
Direct asks, important unanswered threads, commitments.
Focus, Agent Manager, Question Asking, Execution. External sends approval-required.

What was verified

44
unit tests passing
4
sources inspected in live run
4
sources wrote heartbeat records
14
real Hermes faculty agents completed
Quality signal: the first MVP run is competent but too noisy: all 14 faculties surfaced something. The Intelligence Review scores discernment/focus at 2/5 because the next tuning task is stricter suppression and cross-faculty ranking before any Telegram-facing output.

Where to observe/tune

Latest live quality run

/root/personal-orchestrator/runs/2026-05-24T20-36-49.687658-00-00/

  • evidence.md — source evidence packet
  • ingestion_manifest.json — collector/provenance/storage proof
  • faculties/*.md — each faculty judgement
  • AGENT_RUNS.md — faculty run index
  • INTELLIGENCE_REVIEW.md — heuristic review

Main tuning levers

  • Faculty prompts: agents/faculties/*/prompt.md
  • Source policy: state/source_adapter_registry.md
  • Raw data: state/integration_heartbeat.sqlite
  • Quality review: INTELLIGENCE_REVIEW.md
  • Feedback: Agent Manager Approve / Ignore / Edit / Defer logs