AI Coding Agents in Finance

An AI coding agent combines a large language model with a set of tools: reading files, editing them, running shell commands, executing a test suite, and checking the result. It plans a change, makes it, observes whether it worked, and repeats until the task is done or it runs out of its allotted budget. Claude Code and OpenAI Codex are the two tools finance development teams compare most often for this work, each running from a terminal against a developer's own repository.

Agentic software development is the wider practice this points to: a workflow where an agent plans tasks, uses development tools, changes code and checks results with limited step-by-step direction from a developer. A developer no longer types every line. In a finance codebase this covers writing tests, running a migration, or reading an unfamiliar service before a developer touches it, with a human reviewing the result before it ships.

A coding agent's harness is the scaffolding it runs inside: the tools it can call, the rules on what it may touch, and the checks that catch a bad result before it reaches production. Harness engineering is the discipline of designing and tuning that scaffolding, encoding an organization's own standards into how its agents operate. BCG Platinion's work on the agentic software factory names harness engineering as the discipline behind a delivery pipeline where fleets of agents do the coding and engineers define intent and review the result.

Harness engineering is a new discipline, and what counts as good practice is still forming. An agent's failures, and the time it takes to find and fix them, resist the kind of planning a finance team applies to a normal project. The work also competes for priority against features that ship to customers, because a better harness rarely touches a customer directly and can take time to pay off. Visibility is the hardest part: the signal an engineer needs often sits only in one local session or in what a colleague remembers, never written down anywhere a team can search.

Context engineering is the related discipline of deciding what information an agent holds in its working context at each step, separate from how the agent is allowed to act. A bank's platform team usually owns both: the harness that limits what an agent can touch in a regulated codebase, and the context rules that keep an agent from acting on stale or incomplete information about a system it has not fully read.

Agentic software development is an approach where an AI coding agent works toward a stated goal by planning tasks, calling development tools, changing code and checking the result, with limited step-by-step direction from a developer. It differs from autocomplete and from a single-response chat assistant, both of which stop after one suggestion instead of running a loop to a finished result.

An agent harness is the scaffolding an AI coding agent runs inside: which tools it can call, what it is allowed to change, and what checks run on its output before a human sees it. Two agents built on the same underlying model can behave very differently depending on the harness around them, which is why developers compare harnesses, not only models, when they choose a coding agent.

A bank's platform team checks what the agent can read and write, whether its actions run inside the same audit trail as a human developer's commits, and whether a change it proposes still goes through the same review and testing gate before release. The agent's speed does not change what evidence a regulator expects from a production change.