Figure: Google Cloud · ReAct pattern · CC BY 4.0
Code and agents
How coding agents find and change code
Find the right code, bound the tool loop, and verify the patch.
Tony Koo · 5 min read · 17 Oct 2026
Tony Koo / Research
Ideas worth testing.
Essays on machine learning, agentic systems, cloud architecture and decisions under uncertainty. Each one connects a practical example with the original sources.
Independent research notes · Twelve essays · Updated October 2026
Featured essay · AI and agents
When agents need search
Choose the smallest control loop that improves validated outcomes, and be explicit about what branches.
Tony Koo · 4 min read
Read the essay →
Browse the publication
Latest essays
Read across models, systems and decisions, or follow one topic.
18 of 18 items shown
No essays or sources match that search.
Diagram adapted from GitHub Spec Kit workflow · MIT
Delivery and evaluation
From specification to verified release
Turn a reviewed contract and plan into a testable, accountable change.
Tony Koo · 5 min read · 17 Oct 2026
Figure: Google Cloud · MLOps Figure 2 · CC BY 4.0
Machine learning
When an ML model is ready for production
A strong validation score is a start. Deployment needs a baseline, leakage checks and a plan for changing data.
Tony Koo · 3 min read · 15 Oct 2026
Figure: Google Cloud · RAG architecture · CC BY 4.0
Data and retrieval
How data becomes reliable evidence
A good answer needs a trustworthy record, a search method suited to the question and a path back to the source.
Tony Koo · 3 min read · 15 Oct 2026
Figure: Microsoft Azure Architecture Center · Outbox sequence · CC BY 4.0
System design
How to choose a system boundary
Draw the line where state is committed, work is handed off and failures can be retried safely.
Tony Koo · 3 min read · 15 Oct 2026
Figure: Google Cloud · Shared responsibilities · CC BY 4.0
Cloud and operations
What the cloud should own
Managed infrastructure removes some work, but ownership of data, access, recovery and cost still needs a name.
Tony Koo · 3 min read · 15 Oct 2026
Figure: Rbb · Prior, likelihood, posterior · CC BY-SA 3.0
Decisions and money
How to decide under uncertainty
A good process records what was known, what was assumed and what would change your mind.
Tony Koo · 3 min read · 15 Oct 2026
Figure: SEC Investor.gov · Compound-growth chart · SEC reuse policy
Decisions and money
How money compounds and risk interrupts it
Time helps only when a plan can survive its cash needs, costs and the investor’s own behavior.
Tony Koo · 3 min read · 15 Oct 2026
Figure: Google ML Crash Course · Figure 2 · CC BY 4.0
Model foundations
What language models learn
Separate learned weights, task context and verification before trying to fix a weak answer.
Tony Koo · 4 min read · 13 Oct 2026
Figure: Yao et al. · Tree of Thoughts Figure 1 · CC BY 4.0
Architecture and reasoning
When agents need search
Choose the smallest control loop that improves validated outcomes, and be explicit about what branches.
Tony Koo · 4 min read · 13 Oct 2026
Figure: Google Cloud · Human-in-the-loop pattern · CC BY 4.0
Security and governance
Where autonomy stops
Let models propose actions while independent controls decide which effects are allowed.
Tony Koo · 4 min read · 13 Oct 2026
Figure: Google Codelabs · ADK evaluation · CC BY 4.0
Evaluation and operations
How to prove an agent works
Compare a strong simple baseline with an agent under the same conditions, then inspect outcomes and costs.
Tony Koo · 4 min read · 13 Oct 2026
Start here
A path through the ideas
Interesting reads
A small shelf of sources that change a design decision. The notes say why to open each one and what it does not establish.
Attention Is All You Need
The original attention and Transformer architecture. Start with Figures 1–2; remember that today's decoder-only models are not identical.
Building Effective AI Agents
A practical vocabulary for workflows, routing, parallelism and agents. Useful before choosing a multi-agent framework.
Tree of Thoughts
A clear example of branching at intermediate reasoning steps. Its experimental gains are task- and setup-specific.
SYMPHONY
Explores heterogeneous-model planning inside MCTS. Read its benchmarks with the model pool and search budget in view.
OWASP Prompt Injection
A threat catalogue for lower-trust instructions entering an LLM workflow. Pair each threat with an enforceable control.
Demystifying Evals for AI Agents
A practical account of task, environment and grader design. Useful when a fluent trace masks a failed outcome.