Agent intelligence architecture

Making Agents Intelligent

A working hypothesis: agents are not yet intelligent because they are mostly prompt-response systems. To make them intelligent, we need to engineer the rails that let subsystems discern, focus, remember, synthesize, create, learn, and act together.

“These aren't coupled systems — these are intelligently interfacing subsystems to make a proactive agent.”

The goal: move from assistant-as-chatbot to agent-as-operating-system: a system that is omnipresent in the background, quietly compounding memory and capability, surfacing bounded high-leverage cards, and executing safe work under explicit autonomy thresholds.

Foundation primitives
4 → 7

Four current primitives become seven intelligence capabilities when connected through policy, evidence, feedback, and execution.

signalstabsmemoryproposalsskillssynthesisexecution

The intelligence gap is not “LLM smarts”; it is system design.

LLMs provide local reasoning, language, pattern matching, and generation. But intelligence at the agent level requires persistent interfaces between subsystems: what changed, what matters, what is unresolved, what should be remembered, what should happen next, what was learned, and what can safely execute.

Prompt-driven agent

Waits for the user, reasons inside the current context window, loses implicit learning, repeats old mistakes, and confuses activity with progress.

Proactive agent

Continuously senses, filters, remembers, synthesizes, proposes, learns from decisions, and only interrupts when the expected value clears the attention budget.

Intelligent harness

The rails around the LLM evolve: source filters, memory policies, tab prioritization, synthesis patterns, skill libraries, and autonomy thresholds.

Core concern: this intelligence must be built into the agent as capabilities, not as isolated side systems. Signal Radar, Open Tabs, GBrain, proposals, skills, and execution should be interoperable organs of one agentic nervous system.

The current primitives

The foundation is already useful because it compresses personal operating reality into a small vocabulary.

01

Signal Radar

What changed / what might matter.

02

Open Tabs

What is unresolved / needs attention.

03

Memory Capture

What should compound.

04

Proposal Recommendations

What should happen next, and who should do it.

Intelligence emerges from typed interfaces, not tight coupling.

Each primitive should influence the others through explicit conversions, evidence, policy, and feedback. The system should not blindly broadcast every event into every subsystem.

Memorydurable meaning / recall / synthesis
Signalchange detection / salience
Open Tabsattention / unresolved loops
Proposalsnext moves / owner / risk
↕ memory ↔ signal ↔ tabs ↔ proposals ↕
memory + signal + tabsproposals
proposals + memory + signaltabs
tabs + proposals + memorysignal
signal + tabs + proposalsmemory
Design rule: every cross-primitive conversion must improve at least one of attention, memory, action, or learning — otherwise it is noise.

High-signal routing

The same raw integration event can have different destinations depending on what kind of signal it contains.

If the system detects high signal for…
Route to…
Because…
Integration change / external movement
Signal Radar
The event may matter, but does not yet necessarily need action.
Human cognitive load / focus / unresolved commitment
Open Tabs
The user needs an attention object, decision, or follow-up loop.
Current save / future recall
GBrain
The information should compound into durable meaning.
Pattern recognition across memories
Synthesized GBrain
The value is not the fact; it is the connection, drift, contradiction, or repeated shape.
Work, research, recommendations, optimisations, solutions
Proposals
The system should identify an owner, risk, done condition, and proof path.
Procedural memory: current document / future use
Hermes skills
The agent becomes more competent next time, not just more informed.
Safe bounded action
Execution
The system should move from recommendation to verified work when policy allows.

The faculties intelligence must improve into

The important part is the “vs”. Each faculty is a tension the agent must resolve on every tick, then improve through a concrete substrate where learning accumulates.

Discernment

signal vs noise
TowardSignal: meaningful change, repeated pain, leverage, risk, opportunity.
Away fromNoise: novelty, duplicates, low-context feeds, “interesting but not useful”.

Self-improves by turning approvals, ignores, rejects, and missed opportunities into per-source filter rules. Every source gets a learned threshold: what counts as signal here, for Connor, now?

Improves into: per-source filter library → becomes: more discerning.

Focus

now vs later
TowardNow: high-leverage, time-sensitive, cognitively loaded, unblocking.
Away fromLater: safe to batch, evergreen, low urgency, no decision required.

Self-improves through Open Tabs outcomes: what got closed, deferred, ignored, or resurfaced too late. The attention model learns urgency, importance, reversibility, and cognitive load.

Improves into: tab prioritisation logic → becomes: more attentive.

Knowledge

remember vs entropy
TowardRemember: stable facts, preferences, commitments, original thinking, reusable context.
Away fromEntropy: stale fragments, uncited claims, duplicates, context that cannot be retrieved.

Self-improves like Hermes deciding what to store or recall in GBrain: evidence-backed writes, dedupe, expiry, source confidence, and recall success all tune the memory policy.

Improves into: GBrain + memory policy → becomes: more knowledgeable.

Pattern recognition

connect vs isolate
TowardConnect: repeated shapes, contradictions, drift, hidden dependencies, reinforcing loops.
Away fromIsolate: treating each event as a one-off with no memory of prior structure.

Self-improves through synthesis over GBrain: the system learns which connections proved useful, which were spurious, and when a cluster should become a concept, warning, or proposal.

Improves into: synthesis layer over GBrain → becomes: more insightful.

Creativity

generate vs imitate
TowardGenerate: new framings, artifacts, options, diagrams, research angles, product shapes.
Away fromImitate: generic LLM prose, default templates, plausible but unowned synthesis.

Self-improves by saving generation patterns that worked: document forms, visual layouts, prompts, examples, critique loops, and Connor-specific taste constraints.

Improves into: generation template library → becomes: more generative.

Competence

learnt ability vs static ability
TowardLearnt ability: reusable workflows, tested commands, known pitfalls, validation recipes.
Away fromStatic ability: solving from scratch, repeating mistakes, relying only on base-model knowledge.

Self-improves through Hermes skills: every hard-won workflow can become procedural memory with triggers, steps, pitfalls, and verification so the agent is more skillful next time.

Improves into: skills library → becomes: more skillful.

Conscientiousness

proactive vs prompted
TowardProactive: omnipresent background care, safe automatic action, timely escalation.
Away fromPrompted-only: waiting for Connor to remember, ask, triage, or restart the loop.

Self-improves through policy and autonomy thresholds: what can auto-apply, what should be batched, what needs confirmation, and what must never happen without explicit approval.

Improves into: policy / autonomy thresholds → becomes: more autonomous.

Self-improving agent principle: intelligence improves into the harness around the LLM. Filters, tabs, GBrain, synthesis, templates, skills, policies, and execution traces are the persistent organs that make the agent more intelligent over time.

Co-arising faculties: intelligence as simultaneous composition

This is a phenomenological architecture problem: analyze how human intelligence actually feels and functions from the inside, then build agent systems where the faculties coexist, collaborate, and recursively improve together. The point is not seven separate modules. The point is a continuously integrated field of faculties that arise together in every intelligent tick.

A new architectural shift

The architecture has to move beyond isolated tools, dashboards, or pipelines. It must support genuine intelligence faculties that compound and compose with each other: discernment changing focus, focus shaping memory, memory enabling pattern recognition, pattern recognition expanding creativity, creativity creating proposals, competence executing or learning skills, and conscientiousness deciding what should happen next without waiting to be prompted.

1. Integrated perception

Raw integrations and human interactions do not enter a single inbox. They enter a living architecture where multiple faculties inspect the same event at different levels.

discernmentfocusknowledge

2. Shared context, not handoffs

Faculties should not pass a dead packet down a pipeline. They should share evidence, entities, prior decisions, active tabs, memory, capability state, and risk policy.

GBrainOpen TabsSignal Radar

3. Recursive influence

Each faculty changes the operating conditions of the others. Better memory improves signal detection; better signal detection improves memory; better focus changes what proposals are worth making.

memory ↔ signaltabs ↔ proposals

4. Holistic proposal generation

A good proposal is not merely generated. It arises from signal, attention, memory, pattern recognition, creativity, competence, and conscientiousness all informing the next move.

creativecompetentconscientious

5. Continuous background care

The system should become more discerning, focused, knowledgeable, insightful, creative, skillful, and proactive over time — not only during explicit chats.

omnipresentalways workingquiet by default

6. Wholeness as a design constraint

Do not simplify the system into independent scores or siloed automations. The architecture should preserve the whole: continuous, integrated, recursive, evidence-backed, and capable of compounding to an extremely high degree.

wholerecursivecompositional
Non-simplification rule: this is a new arena of problems for agent architecture. The system must mimic the way human intelligence faculties arise together, not reduce intelligence to a checklist of isolated features.

Runtime loop

The agent should run a repeating tick that turns raw input into attention, memory, synthesis, generation, execution, and meta-next-action decisions.

Intake
Discernmentraw input → signal or noise
Attention
Focussignals + tabs → what is now
Memory write
Knowledgewhat to store from this tick
Synthesis
Pattern recognitionwhat connects
Generation
Creativityproduce artifact when useful
Execution
Competenceinvoke skill / adapt
Meta
Conscientiousnesswhat should fire next unprompted

Failure modes to engineer away

The system becomes intelligent by repeatedly converting observed failures into rails, policies, tests, and learned libraries.

Noise amplification

Every source event becomes an interruption, proposal, or tab.

Rail: per-source filters, semantic bundling, quiet compounding, attention budgets.

Memory entropy

The brain fills with duplicated, stale, untrusted, or non-retrievable facts.

Rail: evidence refs, deduplication, confidence, expiry, synthesis, contradiction detection.

False urgency

The agent treats “interesting” as “now.”

Rail: now/later priority model, due dates, reversibility, opportunity cost.

Proposal spam

The user receives too many cards with unclear actions.

Rail: top-N inbox, proposal taxonomy, owner/risk/done condition, rejection learning.

Unbounded autonomy

The agent acts because it can, not because it should.

Rail: autonomy thresholds, allowlists, dry-runs, verification artifacts, rollback paths.

Static competence

The agent solves the same class of problem from scratch every time.

Rail: skill creation, skill patching, validation commands, reusable templates.

Isolated subsystems

Signal Radar, tabs, memory, and proposals become separate dashboards.

Rail: primitive graph, typed conversions, shared evidence ledger, cross-primitive tests.

LLM-shaped hallucinated governance

The model invents policies, capabilities, or completed work.

Rail: capability registry, live status checks, proof paths, source-of-truth files.

Human preference blindness

The system repeats suggestions the user already rejected.

Rail: proposal decision events, semantic suppression, new-evidence reopening rules.

Where Hermes and the LLM sit

The LLM should influence intelligence, but should not be the only place intelligence lives. Hermes is the agent runtime and policy harness; the LLM is a reasoning/generation engine inside that harness.

LLM influence

  • Classification where deterministic rules are insufficient.
  • Synthesis across evidence and memory.
  • Creative proposal generation.
  • Drafting artifacts, plans, and questions.
  • Explaining why a proposal exists.

Harness authority

  • Source registration and evidence storage.
  • Policy gates and autonomy thresholds.
  • Execution allowlists and verification requirements.
  • Memory write rules and expiry.
  • Feedback loops from decisions and outcomes.
Automatic approvals: allow only for low-risk, reversible, source-backed proposal types with explicit evidence refs and a proof artifact. Everything else remains proposal-gated or confirmation-only.

How to build this

Do not build another dashboard. Build the agent’s intelligence substrate: a primitive graph, feedback-trained policies, skillful execution, and a research/execution backlog that compounds.

1. Primitive graph as the substrate

Every event, evidence atom, signal, tab, memory capture, proposal, skill, job, and outcome gets typed edges. The graph is how subsystems interface without becoming tangled.

2. Cross-primitive rules engine

Encode when signal becomes tab, tab becomes proposal, proposal becomes memory, memory becomes signal, and combinations become higher-order recommendations.

3. Autonomy threshold ladder

Define automatic, auto-draft, proposal-gated, confirmation-only, and forbidden classes. Learn thresholds from approvals, rejects, edits, and outcomes.

4. Hermes backlog for research/execution/breakdown

Create a first-class queue of work Hermes should research, execute, decompose, verify, or delegate. Each item needs owner, risk, done condition, and proof path.

5. Self-improvement libraries

Discernment improves through filters; focus through prioritization; knowledge through GBrain policy; pattern recognition through synthesis; creativity through templates; competence through skills; conscientiousness through autonomy policy.

6. Simulation before autonomy

For every integration, run a simulated tick: raw event → signal/noise → loop/memory/proposal/action → approval policy → expected proof. Ship autonomy only after dry-run evidence.

High-level constraint: intelligent subsystems should be interoperable, auditable, and self-improving. If a feature cannot show evidence, explain its conversion, learn from feedback, and prove completed work, it is not part of the intelligence layer yet.