Let agents move quickly inside explicit boundaries.

Jsonic++ is a useful small example of the workflow used across Nift development: the human directs product scope and judgement; agents investigate and execute; contracts and tests preserve what must remain true.

human perception + intent
          ↓
   explicit contracts
          ↓
 agent investigation
          ↓
 implementation + tests
          ↓
 fuzz / sanitizers / consumers
          ↓
        evidence
          ↓
   human checkpoint

Direction is not the same as typing the implementation

An agent can inspect a parser, generate edge cases, edit C++ and run a large test matrix much faster than a human can do each mechanical step. The high-leverage human work is deciding what the parser is supposed to mean: whether duplicate keys are accepted, whether numeric coercion belongs, whether an apparently convenient feature would permanently enlarge the API, and what evidence is sufficient before a change is accepted.

Start from a falsifiable guarantee

“Improve Unicode support” is too vague. “Accept valid escaped surrogate pairs, reject lone high/low surrogates, preserve the decoded value through serialization, and keep useful line/column failures” gives an agent something it can attack. Good contracts create an adversarial task rather than a code-generation task.

Project memory, not model memory

Important decisions belong in tests, handovers and source—not only in a conversation. A new agent should be able to recover why duplicate keys are rejected, why numbers are represented as double, or why vendored copies must match without relying on someone remembering a chat transcript.

This also makes work resumable. If one session ends after discovering a malformed-number family, the next session should inherit a reproducer, an explicit expectation and the current validation state—not just a claim that “number parsing was being improved.”

Agents should try to disprove the patch

The most useful second pass is often adversarial: search for siblings of the fixed case, stress nested/empty/boundary forms, run sanitizers, compare parser and serializer behavior, and check downstream consumers. The goal is not to produce more tests for their own sake; it is to discover whether the stated contract has a larger failure surface than the first reproducer suggested.

Perception matters

Agents can implement a well-framed parser change extremely quickly. Human direction remains high leverage when the real question is whether the proposed feature belongs in Jsonic++ at all, whether the test actually captures the intended guarantee, or whether complexity is growing faster than value.

The checkpoint question

For practical agent prompts, see AI assistants. For parser-specific development habits, see AI development.