This site
Artificial Quirks is a working notebook on how agent systems actually behave in production. It covers real runs and real failures, not demos or vendor decks.
The industry is shipping recursive loops, multi-model fleets, and "just add RAG" stacks at high speed. The interesting problems are quieter: which knobs actually move cost, when a prettier metric hides a failed job, and how to keep classical software discipline when the runtime is a language model.
What you'll find here
- Agent runtime design: cell budgets, history working sets, strategy presets instead of model-name
ifs. - Evaluation that measures the job: completion and artefacts alongside cells per turn and token burn.
- Security and sandboxing: what an agent can touch, and the audits that show it.
- Platform notes: architecture trade-offs for GenAI platforms, risk vectors, and operational honesty.
Claims stay labelled: observation, inference, provisional. Numbers come with horizons, stop conditions, and limits.
The same methods are available as a fixed-scope review of your agent system.
Current series
Recursive language-model loops: three notes on cells, history, and named strategies.
- 1 When Smaller Cells Make Agents Worse
Prompt-only “smaller cells” improved aesthetics and missed completion under a short step budget.
- 2 RLM Is Not Automatically Token-Efficient
History policy plateaus input tokens; cheaper runs still need deliverable checks.
- 3 Strategies, Not Model Ifs
Named YAML strategies beat model-id branches in the agent runtime.
All posts
Topics
-
Agent runtimes
Recursive loops, code cells, tool budgets, and the policies that keep long-horizon agents from burning the ledger.
-
Evaluation & honesty
Measure the deliverable—not only proxy aesthetics. Pair tokens, cells/turn, and step use with FINAL and artefacts.
-
Strategy over model ifs
Named execution strategies, profile bindings, and task overrides—without hardcoding vendor strings in the engine.
-
Security & sandboxing
What an agent can actually touch: isolation, network egress, credentials, and audits of real runs.
Work with me
Agent Runtime & Evaluation Review (2 weeks, fixed fee). An independent, evidence-based review of an agent system before you scale it: what it can touch, how it fails, what it really costs, and whether your evaluation would notice.
Petar Kabashki
Independent advisor · agent systems & production GenAI
20+ years in software, much of it in banking
Recent posts
- Your Green CI Check May Have Run Zero Steps
A passing check is a claim about execution, not proof of it. Four distinct failure modes in one pipeline turned absence-of-failure into a success signal.
13 Oct 2026
- The Human Approval Path That Never Ran
We built signed approvals, attested principals and dual sign-off. Across three weeks and 542 runs, all 46 approvals were recorded as 'system'.
8 Oct 2026
- Why Build an Agent Harness?
An agent harness turns coding-agent activity into a bounded, inspectable, and repeatable engineering process rather than a stream of opaque terminal outcomes.
5 Aug 2026
- Zero-Spend Concurrency Auditing for Agent Sandboxes
A mock-driven diagnostic method for exercising sandbox resource cleanup without live-model spend.
29 Jul 2026
- Catching Hallucinated Paths in RLM Trajectories
How pre-finalization trajectory hooks catch ungrounded file paths and trigger self-correction turns before final answer submission.
23 Jul 2026
- Grounding Is Not Proof: 100% Code, 0% Plan Quality
Why a 100% task completion rate does not guarantee plan quality, and how pre-finalization checks prevent hallucinated path claims.
23 Jul 2026
- Inside the Sandbox: Unconstrained Shell Execution
Exploring why allowing agents arbitrary host shell access leads to non-deterministic failures and why strict Docker isolation is mandatory for production.
23 Jul 2026
- The Security Cost of Autonomy: Auditing 29 Runs
Empirical security audit results from 29 benchmark study runs evaluating container mount points, path traversal attempts, and write-tool boundaries.
23 Jul 2026