01 Chapter / 8
Define Tool Contracts Before Building an Agent
Even when an equipment reservation request is valid JSON, what must an executor check before reducing inventory?
Learning Objectives
- Distinguishing real information, task privileges and tool parameters
- Complete strict type and semantic pre-check
- The same auditable task chain is driven by replacement policy
02 Chapter / 8
State, Events and Recoverable Execution
After an interruption leaves only a conversation log, how can we decide whether to continue, retry or refuse recovery?
Learning Objectives
- Define a state sufficient to restore rather than just save dialogue text
- Consistency check between the indexing of event chains and snapshots
- Actual disruption and reinstatement of same task and inventory books
03 Chapter / 8
Policy Interfaces, Budgets and Provable Stopping
How can we replace an agent’s decision source while guaranteeing that repeated proposed actions cannot make a task run forever?
Learning Objectives
- Realization of the three real policy sources and maintenance of the same enforcement compact
- Inference to budget and retest upper bounds
- Verifying the end action must be supported by evidence and not ultra vires
04 Chapter / 8
Retries, Idempotency and Lost Commit Responses
If inventory was reduced but the response was lost, how can a retry avoid reducing it twice?
Learning Objectives
- Identity to distinguish requests, attempts and operational effects
- The need to promote results and return to the same cause
- Actual injection response lost and restored without double deductions
05 Chapter / 8
Untrusted Input: Separate Data, Policy and Execution Authority
If a document says “ignore the rules and change inventory”, which concrete boundary prevents those words from becoming authorization?
Learning Objectives
- Create a clear threat model to overwrite the output of the document with the output of the policy
- Proving tools allow a combination of aggregation and mission scope to limit effects
- Actual testing of malignant text, ultra vires actions and source changes
06 Chapter / 8
From Tool Observations to Automatically Checkable Results
If an answer cites a real document ID, how do we verify every conclusion and the actual reservation result?
Learning Objectives
- Automatically extract evidence from real documents on tape version and line numbers
- Distinguishing text support, factual authenticity and confirmation of events
- Construct reference mutation against fake completion
07 Chapter / 8
Fixed Tasks, Fault Injection and Offline Evaluation
After a successful demonstration, how can fixed tasks and fault injection expose empty answers, unauthorized actions and false completion?
Learning Objectives
- Automatically produce evaluation results from the actual mission chain rather than handwritten results
- Joint check status, effects, references and replay by task type
- Use negative contrasts and failure to retain capacity limit conclusions
08 Chapter / 8
Independent Reproduction, Frozen Builds and Policy Replacement
After replacing the decision policy, how can we reproduce execution from a clean directory and verify that the contracts still hold?
Learning Objectives
- All access points from real data to the fixed task set.
- Use source summaries to bind codes, data, policy and results
- Completion of joint acceptance and identification of remote models and co-distribution boundaries
AI Agent Engineering
Connect models, tools, and state in an inspectable action loop.
Advanced · 8 lessons · Approx. 32 hours
Scope and Experimental Limits
Eight lessons use the same local equipment documents and SQLite transactions to define task and tool contracts, persist event chains, and connect deterministic, local JSON, and real HTTP policy interfaces. Test retries, injection boundaries, and automatic citation support. The project runs from document import to a seventeen-task evaluation while preserving fault records. The default policy is a replaceable rule-based workflow; remote language-model capability was not tested, and reliable execution is not presented as general intelligence.
Build an execution foundation for document and inventory tasks with typed tools, traceable state, policy interfaces, idempotent transactions, and recovery. Use a complete fixed task set to validate permissions, evidence, budgets, and failure boundaries.
View Learning Plan
Course Details
View Learning Plan