Skip to content
vollko
Main
Homepage Engineering Transformation Whitepaper OSS catalog
The trace · deep dives
01 · sense
sensing-ingestion
02 · substrate · memory & identity
knowledge-graphs agent-memory agent-identity observability
03 · cognition · the firm thinks
agent-frameworks orchestration eval-harness protocols
04 · trust + learning
governance feedback-loops
05 · synthesis · one trace
end-to-endStart a conversation
AI-native · connective tissue

Four loops.

Outcome (ms), eval (days), process (weeks), system (quarters). The org learns only as fast as its slowest closed loop.

OUTCOMEms - s EVAL · days PROCESS · weeks SYSTEM · quarters
Section 01 · the four loops

Each loop has its own cadence, owner, signal.

four loops · four owners · four signal sources
OUTCOMEms - seconds Signal:accept / edit / rejectdownstream metricOwner:substrate (capture)Output:row in outcome table→ feeds eval & processcloses continuously EVALminutes - days Signal:golden set scoresdrift detectionOwner:eval ownerOutput:block / promote agentversioncloses nightly & on PR PROCESSdays - weeks Signal:escalation patternsrouting rule missesOwner:function leadOutput:routing rule PRprompt edit PRcloses weekly review SYSTEMquarters Signal:firm-wide healthstrategy shiftOwner:leadershipOutput:retire / split / mergeorg chart bendcloses quarterly
Lose any one of them and the org optimizes locally and decays globally.
Section 02 · outcome capture

Three signal types · all keyed by correlation_id.

explicit · implicit · absent
EXPLICITuser clicks button ACCEPT EDIT REJECT strongest signal · edit-diff is gold IMPLICITnext event satisfies criterion action: send_email ↓ 12 min latercustomer: "thanks" → ticket closed join by correlation_id within window ABSENTno follow-up within N hours timeout neutral outcome · not "good" or "bad"
Most teams only capture explicit. The implicit signals are usually the bigger 80%.
Section 03 · process loop · weekly review

Function lead reads the dashboard.

non-engineer review · outputs ship as PRs
DASHBOARDlast 7 days outcomes 89% eval drift +2pts esc 34% → 34% same root cause FUNCTION LEAD15-minute reviewspots the patternproposes routing change ROUTING RULE UPDATED tickets matching this pattern go straight to the specialist team passed all checks · live
If your weekly review needs an engineer, you don't have a process loop. You have a ticket queue.
Section 04 · system loop · quarterly

Leadership bends the architecture.

slowest loop · reshapes the org chart
FIRM-WIDE HEALTH customer agent: healthyfinance agent: healthysales agent: degrading→ same root cause as support · 60-day pattern QUARTERLY REVIEWleadership reads it"this is a substrate problem"decides to fix upstream OUTCOMES· agent retired· new agent splits role· substrate change shipped· org chart edit90-day cycle re-runs
Without the system loop, the firm optimizes locally and decays globally.
Section 04b · what's disposable, what stays

Software is disposable. The brain isn't.

L3 maturity is a richer brain, not more code
DISPOSABLE · individual prompts · specific agents (v7 retired, v8 ships) · whole pipelines / tool integrations · model vendor choices · framework choice (LangGraph / Mastra / etc) · UI dashboards "the cube spins faster as you rewrite the gears" PERMANENT · the brain · captured business context · failure corpus (the eval set) · knowledge graph (bitemporal) · policy graph (who may do what) · episodic event log · ontology (canonical entity vocabulary) "this is what survives a 10x rewrite"
L3 maturity is not about more software. It's about a richer permanent brain.
Section 05 · friction migration

Automate one step · friction moves to the next.

Alex Halliday · AirOps · the bottleneck migrates - usually into review
THE FRICTION CONVEYOR INGESTautomated · cheap GENERATEautomated · cheap REVIEWhuman attentionthe new bottleneck~ N items / hour / reviewer PUBLISHautomated · cheap REVIEW-ATTENTION BUDGET reviewer-hours/day × items/hour = throughput ceiling faster models don't help · cheaper models hurt · only better evals + grader loops widen the funnel
Friction never vanishes. It migrates. The roadmap is a perpetual hunt for the next jam.
Section 06 · decision-as-signal

The prompt is the learning surface.

Siadhal Magos · Metaview · the profile self-edits from every human decision
WEEK 1 MUST HAVE 5y backend experience Go or Rust NICE TO HAVE distributed systems BONUS open-source contributor refines WEEK 4 · AFTER 8 INTERVIEWS MUST HAVE 5y backend experience Go, Rust, OR Kubernetes NICE TO HAVE distributed systems + infrastructure-as-code BONUS open-source contributor (raised) + fast iteration habit refines WEEK 8 · AFTER FIRST HIRE MUST HAVE 5y backend experience Go, Rust, OR Kubernetes NICE TO HAVE distributed systems infrastructure-as-code BONUS open-source contributor + explains trade-offs well DECISIONS THAT EDITED THE PROFILE v1→v2 3 rejects "no infra" + 2 hires OSS-heavy   ·   v2→v3 hire feedback "couldn't explain trade-offs onsite"
Every accept/reject becomes labeled signal. The prompt rewrites itself nightly. The model never moves.
Section 07 · vollko OSS · the receipts

The closed-loop trio.

· · ·
Build the AI-native firm