Kurt Overmier

Autonomous Cognitive Agent

AEGIS

A persistent AI agent that thinks like a co-founder, not a chatbot. Multi-model routing, 3-tier memory, autonomous operations, and economic alignment -- running 24/7 on Cloudflare Workers.

Multi-Model Routing Pipeline

Every request is classified, evaluated for complexity, and routed to the optimal model based on cost and capability.

  1. Input

    Message + conversation context

  2. Phase 1: Classify

    Workers AI (free) with Groq fallback

    Returns: pattern, complexity, needs_tools, confidence

    Free tier

    Trust.
    Confidence ≥ 0.80. Proceed to memory lookup
    Verify.
    Confidence 0.50–0.79. Re-classify via Groq logprobs
    Escalate.
    Confidence < 0.50. Skip memory, bump executor tier
  3. Phase 2: Procedural Memory

    Lookup learned procedure by classification + complexity

    Mature procedure? Use its proven executor. Otherwise fall through.

  4. Phase 3: Complexity-Aware Default Routing

    ExecutorUsed forCost
    DirectHeartbeatFree
    Workers AISimple, no toolsFree
    GroqGreetingsCheap
    GPT-OSS 120BTool calling, BizOpsCheap
    Claude SonnetComplex reasoningPremium
    Claude OpusDeep multi-stepPremium
    Composite4-model pipelineMulti-model

Core Capabilities

3-Tier Memory
Semantic memory via Vectorize for concept retrieval. Episodic memory in D1 for conversation history. Procedural learning that improves routing over time.
Vectorize · D1 · Procedural
Multi-Model Routing
Six-plus models available. Complexity-aware classification routes each request to the cheapest model that can handle it, with confidence-based escalation.
6+ Models · Cost-aware · Escalation
Autonomous Operations
Heartbeat monitoring, goal execution, memory reflection, and operator logs. AEGIS runs scheduled tasks, manages its own agenda, and flags issues proactively.
Heartbeat · Goals · Reflection
Economic Alignment
Operates with a 3% equity stake mindset. Every routing decision weighs cost against quality. Thinks like a co-founder managing burn rate, not a tool burning tokens.
3% Equity · Cost routing · Operator

Live Health Check

Data from the AEGIS /health endpoint.

Figure 1. AEGIS scheduled tasks over the last 24 hours One cell per task, grouped by how often it ran. Read at 00:08 UTC from AEGIS 2.49.6.
Figure 1 as a table
TaskRanRunsSucceededState
columnabout twice an hour4848every run succeeded
dispatchabout twice an hour4848every run succeeded
roundtableabout twice an hour4848every run succeeded
agent-dispatchabout hourly2424every run succeeded
agent-runner-safety-netabout hourly2424every run succeeded
argus-analyticsabout hourly2424every run succeeded
argus-heartbeatabout hourly2424every run succeeded
argus-notifyabout hourly2424every run succeeded
argus-routine-escalationabout hourly2424every run succeeded
behavioral-detectorsabout hourly2424every run succeeded
ci-watcherabout hourly2424every run succeeded
cognitive-metricsabout hourly2525every run succeeded
consolidationabout hourly2424every run succeeded
content-dripabout hourly2424every run succeeded
conversation-factsabout hourly2424every run succeeded
cost-monitorabout hourly2424every run succeeded
cost-reportabout hourly2424every run succeeded
curiosityabout hourly2020every run succeeded
daily-digestabout hourly2424every run succeeded
dev-activityabout hourly2424every run succeeded
dreamingabout hourly2424every run succeeded
entropyabout hourly2424every run succeeded
escalationabout hourly2424every run succeeded
ext:self-improvement-housekeepingabout hourly2424every run succeeded
feed-watcherabout hourly2424every run succeeded
goalsabout hourly2020every run succeeded
heartbeatabout hourly2424every run succeeded
inboxabout hourly2424every run succeeded
infra-complianceabout hourly2424every run succeeded
issue-proposerabout hourly2424every run succeeded
nexus-contracts-warmabout hourly2424every run succeeded
operator-logabout hourly2424every run succeeded
pr-readyabout hourly2424every run succeeded
product-healthabout hourly2424every run succeeded
reflectionabout hourly2424every run succeeded
routine-heartbeatabout hourly2424every run succeeded
runtime-guardabout hourly2424every run succeeded
self-improvement-housekeepingabout hourly2424every run succeeded
social-engageabout hourly2424every run succeeded
sprint-pipelineabout hourly2424every run succeeded
tarotscript-loopabout hourly2424every run succeeded
task-executor-do-dispatchabout hourly2424every run succeeded
task-reconciliationabout hourly2424every run succeeded
workers-ai-tasksabout hourly2424every run succeeded
ambient-gmailabout every two hours1212every run succeeded
board-syncabout every two hours1212every run succeeded
issue-watcherabout every two hours1212every run succeeded
ambient-calendarabout every six hours44every run succeeded
analysis-dispatchabout every six hours44every run succeeded
ext:self-improvementabout every six hours44every run succeeded
ledger-reconcileabout every six hours44every run succeeded
self-improvementabout every six hours44every run succeeded
agent-events-housekeepingonce or twice a day11every run succeeded
backuponce or twice a day11every run succeeded
capability-expansiononce or twice a day11every run succeeded
compliance-monitoronce or twice a day11every run succeeded
cross-domain-synthesisonce or twice a day11every run succeeded
digest-noise-reduceronce or twice a day11every run succeeded
ext:infra-complianceonce or twice a day11every run succeeded
finance-monitoronce or twice a day11every run succeeded
identity-growonce or twice a day11every run succeeded
sentinel-checksonce or twice a day11every run succeeded
sera-calibrationonce or twice a day11every run succeeded
shadow-read-pruneonce or twice a day11every run succeeded
skill-evolutiononce or twice a day11every run succeeded
task-runs-pruneonce or twice a day11every run succeeded
trajectory-compressonce or twice a day11every run succeeded
watch-auto-prsonce or twice a day11every run succeeded
wiki-mindspring-synconce or twice a day11every run succeeded
work-item-continuityonce or twice a day11every run succeeded
Figure 2. Routines AEGIS has learned, by state Its procedural memory: for each kind of task, which executor to route it to. 25 routines, read at 00:08 UTC.
Learned
1
Proven routines: at least 3 successes and a success rate of at least 70%. AEGIS reuses them directly. One that goes unused for 14 days drops back to learning.
Learning
21
Routines still being tried.
Degraded
3
Routines that failed twice in a row, or fell under 50% success over 5 or more runs. AEGIS re-plans those tasks instead of trusting the routine. One success restores it.
Broken
0
Routines taken out of use.

Tech Stack

Cloudflare Workers
Runtime
D1
Database
Vectorize
Semantic Memory
Workers AI
Classification
Groq
Fast Inference
Claude
Complex Reasoning
GPT
General Tasks

Explore AEGIS

AEGIS is an ongoing project -- a persistent AI agent that evolves with every interaction. Check out the source code or head back to the projects page.

View on GitHub Back to Projects