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 · substrate

Workflows that survive crashes.

Temporal, Restate, DBOS. Every LLM call is an activity, every approval a signal, every retry a free win.

1start 2retrieve 3llm.draft await_signalhuman approval 4tool.send 5end CRASH at step 3worker dies mid-LLM call DURABLE STATE STOREstep 1 ✓ · step 2 ✓ · step 3 (in flight)→ on worker restart: resume from step 3 with same idempotency key
Section 01 · Model · Harness · Orchestrator

Three concentric layers.

Alan Walsh · the AI Automators · the layer almost everyone skips is the outermost one
ORCHESTRATOR coordinates teams + humans HARNESS supervises every run TOOLS does the work MODEL thinks WHAT MOST TEAMS MISS Most teams build onlythe inner two rings.Production needs all four - so a bad run doesn't ship.
A capable model in a thin harness is a demo. Production is the outer two rings.
Section 02 · the Ralph loop

Spawn. Run. Sensor. Repeat.

an outer-harness pattern · respawn fresh sessions until the sensor passes
SPAWN fresh session no prior context RUN agent attempts task inner harness drives SENSORS det. + inf. verdict linter · LLM-judge PASS RESPAWN try N + 1 fresh context · no contamination
A spawn-respawn loop bounded by sensors. Each iteration is isolated; no shared scratchpad to corrupt.
Section 03 · planner · executor · DSL

The plan is an artifact. Not prose.

Andreas Stuhlmüller · Elicit · LLM writes the plan; interpreter runs it
PLANNERLLMwrites the plan THE PLAN · a recipe, not a paragraph 1. search papers 2. keep only the recent ones 3. summarize each 4. rank by citations fixed shape · no infinite loops · cost-capped CHECKER GATE ✓ types✓ cost estimate✓ policy (Cedar) EXECUTOR deterministic interp no LLM in this loop search()summarize()rank() VERIFIABILITY SPECTRUM free prompt json tools DSL plan code-as-action ← flexible, harder to check checkable, less flexible →
The model writes a plan you can type-check, cost-estimate, policy-stamp - before any tool runs.
Section 04 · plan · execute · observe · backtrack

Backtrack is the power.

Kay Zhu · Genspark · workflows accumulate errors; agents recover from them
AGENT LOOP PLAN EXECUTE OBSERVE BACKTRACK recovery seam BRITTLE WORKFLOW step step step × break tiny errors pile up · eventually one breaks the chain errors absorbed at each step
Workflows accumulate errors. Agents recover - by re-observing reality, not the script.
Section 05 · the difference

async/await vs durable execution.

same recipe · very different crash behavior
THE OLD WAY 1. fetch context 2. draft reply 3. wait for approval - CRASH 4. send email - never happens lose everything · start overduplicate emails · lost approvals DURABLE WORKFLOW 1. fetch context 2. draft reply 3. wait for approval - CRASH 3. resumes from here 4. send email - once crashes resume from the last completed stepnever sends twice · approval stays in queue
Almost the same code. Wildly different production behavior.
Section 06 · activities

Every external call is an activity.

workflow = pure orchestration · activities = the side-effects
WORKFLOWplans the work 1. fetch context 2. draft reply 3. wait for approval 4. send email every step saved · replayable FETCHretries up to 5× DRAFTretries 3× · never duplicates SENDretries 3× · sends once only vector DB / API LLM provider mail gateway
The workflow body never touches the outside world directly. That's what makes it replayable.
Section 07 · human-in-the-loop

wait_for_signal - a 24-hour pause.

workflow sleeps free · resumes when human signals
t=0draft ready PAUSED · 4h 12m human approves resume send done worker memory: 0 · cost: 0 · the workflow is just a row in the store waiting for an event
No polling. No cron. The workflow costs zero while it waits.
Section 08 · idempotency

Same key, same outcome.

deterministic key per side-effect
send_email key = sha256( ticket_id + draft_sha + recipient) KV STORE checkSETNX key 1 EX 24h→ already exists? send skip customer sees 1 email even after 5 retries
Without idempotency, retries are dangerous. With it, retries are free.
Section 09 · tools 2026

Where the OSS picks land.

Postgres-native ↑↓ dedicated infra · OSS ←→ SaaS-only
SaaS-onlyOSSPostgres-native (low-ops)dedicated infra DBOSlibrary + Postgres · zero ops Restatedurable serverless · Postgres or own Temporal OSSthe gold standard · self-host HatchetAI-DAG · OSS Trigger.dev v4OSS core · TS-first Inngest OSS Temporal Cloudpaid managed AWS Step FunctionsAWS lock-in agent-task-protocolvollko · engine-agnostic
DBOS for new teams already on Postgres. Temporal OSS when you outgrow it. Skip AWS Step Functions for AI workloads.
Section 09b · agent-task-protocol

A portable contract for what a task is.

one shape every framework can speak
EVERY FRAMEWORK REINVENTS A TASK LangGraph Temporal MCP Task OpenAI Agents SDK Inngest five tools, five vocabularies · nothing moves between them ONE SHARED VOCABULARY IDENTITY who, what, version STATUS where in its lifecycle BUDGETS time, cost, tokens RETRIES how many, how long CHECKPOINTS saved progress APPROVALS human gates passed frameworks adapt · tasks become portable
Six things every task needs - finally portable across tools.
Section 10 · vollko OSS

The primitives.

· · ·
Build the AI-native firm