Intent
The system knows what outcome it is serving.
Paweł Szóstakowski / AI automation & agentic systems
Web systems that can act - and still answer for what they do
Production AI workflows, multi-domain platforms, autonomous narratives, and agent-connected infrastructure.
I design and build systems that observe state, coordinate agents, make bounded decisions, and remain open to human intervention. This site documents the architecture, operating principles, and working projects behind them.
Engineering model
Six ideas, one operating model
Six responsibilities that have to work together in a governed autonomous system.
010203040506
Standing systems
The web as a system that keeps observing and deciding after deployment.
Explore the thread ↗Bounded agency
Agents act inside explicit authority, review gates, and reversible states.
Explore the thread ↗Narrative state
Content composed from goals, evidence, relationships, and current context.
Explore the thread ↗Adaptive experience
Interfaces that change with state without concealing why they changed.
Explore the thread ↗Orchestration
Specialized roles cooperating through contracts instead of one opaque model.
Explore the thread ↗Operational evidence
Build logs, failure modes, overrides, and what survived contact with reality.
Explore the thread ↗Controlled publishing loop
Generation is one chamber, not the whole pipeline
The essential distinction is between producing a candidate and authorizing a public state change.
- 01
Signal
A request, observation, or dated source enters with provenance.
- 02
SourcePack
Evidence is gathered before a model is asked to compose.
- 03
Candidate
Generation produces a reviewable proposal, never the final truth.
- 04
Audit
Claims, links, scope, voice, and risk are checked explicitly.
- 05
Promote
A human or policy gate changes publication state.
- 06
Observe
The published result remains measurable, reversible, and correctable.
Minimum viable governance
Autonomy needs a contract
Quality is not a property of the model alone. It emerges from constraints, state, evidence, and a credible way to intervene.
Authority
Every actor has a visible boundary on what it may change.
Evidence
Decisions can be traced back to sources and system state.
Intervention
A person can inspect, stop, or replace a decision.
Reversal
Publication is a state transition that can be undone safely.
Learning
Failures become constraints for the next run, not hidden anecdotes.
Evidence from working systems
What building these systems changed in my approach
Working systems turned broad design principles into concrete engineering requirements: durable state, bounded authority, traceable evidence, explicit publication states, human intervention, and safe reversal.
Explore the case studiesSelected reading
Engineering notes grounded in working systems
Field note
Generation Is Not Publication
Why an autonomous content pipeline still needs human override, observable decisions, and a hard boundary between producing a draft and making it public.
Field note
What Broke While Building an Autonomic Publishing System
A build log about the assumptions that failed, the problems hidden by technically correct pages, and what a multi-domain experiment taught me about autonomy.
Field note
The Autonomous Future: From Automation to Organisms
Automation repeats a designed path; autonomy adapts under intent and constraints. The organism metaphor is useful only if it does not hide ownership, infrastructure, or human...
Field note
limina.li: A Working Example of an Autonomic Narrative System
A dated, public-evidence case study of limina.li: what its expanding narrative graph demonstrates, what can be inspected today, and what the experiment does not yet prove.
Field note
Markets, Feeds, and the Cost of Letting Algorithms Decide
Trading algorithms and recommender systems are not one category of autonomy, but both reveal how objectives, feedback loops, scale, and weak intervention can turn local success into...
Latest state
Recently published
New case studies and engineering notes document how the systems evolve in practice.
- 01Generation Is Not PublicationRead ↗
- 02What Broke While Building an Autonomic Publishing SystemRead ↗
- 03The Autonomous Future: From Automation to OrganismsRead ↗
- 04limina.li: A Working Example of an Autonomic Narrative SystemRead ↗
- 05Markets, Feeds, and the Cost of Letting Algorithms DecideRead ↗
- 06Edge, Home, and Logistics: Autonomy Beyond the BrowserRead ↗