Module Deep Dives
Follow the mechanisms that take an agent from an input through research, tool execution, stored results, memory and final publication. Translations are published separately and retain the original article identity.
New custom-expert case study: A Custom Expert Should Be More Than a Prompt Field follows account revisions, frozen Run identities, Child method inputs, History attribution, and lessons from real integration defects. It also compares the design with Octop's persistent expert instances.
A08 deep dive: Durable control events separates original business receipts, claim/fence validity and delivery acknowledgment.
A30 deep dive: Provider capacity leases after HTTP 200 covers shared endpoint queues, streaming terminals and unknown occupancy.
A10 deep dive: Why a Unified Model API Still Needs Shared Admission now also explains A32 catalog/route eligibility. The separate A30 essay owns the full shared-capacity lifecycle.
A35 deep dive: Prompt Injection Is More Than Malicious Text explains producer-bound instruction identity, final SDK projection checks, and local effect authority. The evidence is structural; real-model resistance remains unqualified.
New mechanism deep dives: Durable human response adoption explains HIO request/response identity and the actual research consumer; temporal semantics distinguishes user days, exchange sessions and data-as-of facts; Shadow Credits tracks physical model attempts with explicit unknown usage. Their local implementation evidence is not production qualification.
01 / Module Deep Dives
Concrete mechanisms from input through execution to delivery.
Research, Reading and Computing
How Entry, Web, PDF, subagents and isolated computation cooperate.
- Entry: Why the Entry Point Became Simpler (Chinese original)
- Web Research: From Search to Deep Reading (Chinese original)
- When an Agent Says Today: User Time and Market Sessions
- PDF: From Text to Evidence
- Subagents: Isolation and Specialized Collaboration
- A Custom Expert Should Be More Than a Prompt Field
- Adaptive Research and Python
- Python: From Tool to Computational Workbench
- Connecting Portfolio Data Through a Read-Only MCP (Chinese original)
- Controlled Refresh for Stale Data
Context, Results and Knowledge
What the model sees and which evidence it may read.
Runtime, Extensions and Observability
How capabilities are discovered, authorized, executed and observed.
- Provider: Why a Unified API Still Needs Shared Admission
- Why the Next Model Request Still Queues After HTTP 200
- Extension Host: From Feature Phone to Smartphone
- MCP Host: From Discovery to Usable Tools
- Tool Surface: Why the Model Cannot See an Available Tool
- Prompt Injection Is More Than Malicious Text: An Agent Content Trust Boundary
- Multi-Ecosystem Hooks: Claude, Pi and Codex
- Observing the Tool-to-Model Input Lifecycle
- Run Budget and Constraint Governance
- Agent Usage: Requests, Tokens and Shadow Credits
- Why an Agent Cannot Execute a Stop Command Immediately
- Scheduled Research Beyond a Conversation
- Langfuse: From Traces to Trustworthy Observability
Answer Delivery and Recovery
An answer candidate, durable effects and delivery are separate facts.