Engineering Autonomous Agents
An agent’s capabilities depend on its data, execution loop, state, and feedback. Design them as one system.
A coding agent is asked to replace a deprecated library across a repository. It finds the imports and writes a patch. Completing the job still requires a dependency map, a test environment, permission to change files, and a way to recover if the task stops halfway through.
The model participates in this system. Its surroundings determine whether an investigation can become completed work.
Give the task an observable outcome#
“Modernize this repository” leaves success ambiguous. A clearer task asks the agent to replace a specific dependency, preserve public behavior, and produce a patch with passing tests.
Specify the target version, affected packages, compatibility requirements, and permitted changes. The agent can choose its approach while the system retains a concrete definition of completion.
Put repeated mechanics in code#
Authentication, pagination, retries, and unit conversion should not become fresh reasoning problems on every turn. Collect data through ordinary code and expose a stable interface for exploration.
The same applies to execution. A tool result should distinguish success, failure, and an unknown outcome. If an action times out, check whether it happened before retrying it. Otherwise a network interruption can become a duplicate operation.
Preserve progress across turns#
A long investigation needs more than a transcript. Keep references to evidence, current hypotheses, completed actions, and open questions in explicit state.
After a restart, the agent should know which packages it migrated, which tests failed, and which design question remains open. That continuity reduces repeated work and makes handoffs inspectable.
Close the loop with feedback#
Check the resulting artifact. Does the patch build in a clean environment? Do integration tests exercise the changed behavior? Does the new dependency work on every supported runtime? The answers separate a plausible patch from a completed migration.
Evaluate the model together with its data interfaces, execution loop, and state handling. Changing any one of them can change the result. That is the unit of engineering: an agent operating inside a complete system.