Coding Agents
Coding agents are AI systems, built on large language models with tools, that plan and carry out software-development tasks with a degree of autonomy — reading a codebase, changing code across files, running tests, fixing what fails, and preparing a pull request for a person to review — rather than only suggesting the next line. They became the fastest-adopted category of enterprise AI and the subject of the largest AI application deals of 2025.
- Also known as
- AI coding agents, Autonomous coding agents, AI software engineers, Agentic coding
- Used at
- feature development, bug fixes, test writing, refactoring and migrations, code review, documentation, and the operational scripts around production; in a fintech's engineering organisation, under review and change-control processes that the agent's output must pass like any other
- Binding
- no rule on the tool; the outputs are subject to the company's change-control, security and operational-resilience obligations (DORA's ICT change management, the security review of code that touches money); licensing questions on training data and generated code; data terms on what the agent sees (the codebase, secrets, customer data in fixtures)
- Typical capabilities
- understanding a repository, planning a multi-step change, editing many files, running builds and tests in a sandbox, iterating on failures, using tools (search, terminals, browsers, issue trackers), and producing a reviewable pull request with an explanation; typical limits: mistakes that look plausible, drift from the intent on long tasks, security errors, and the need for a person who understands the code to review
- Typical economics
- subscriptions per developer plus usage priced by token or by task; reported productivity gains vary widely (from single digits to multiples, depending on task and measurement); the vendors' revenues grew from near zero to hundreds of millions in two years; the labs and the IDE companies compete for the same developers
- Parties
- the labs (Anthropic's Claude Code, OpenAI's Codex, Google's offerings), the IDE and tool companies (Cursor, Windsurf, GitHub Copilot, Replit, Devin's maker), the enterprises adopting them, the security and platform teams governing them, and the developers whose work changes
How a coding agent works in practice
An engineer at a payments company assigns an agent a ticket: add a field to the payout API, validate it, store it, expose it in the dashboard, and cover it with tests. The agent reads the relevant code, proposes a plan, edits the API schema, the validation, the migration, the dashboard component and the tests, runs the test suite in a sandbox, fixes two failures it caused, and opens a pull request with a summary of the changes and the test results. The engineer reviews it — reading the migration carefully, because it touches the payouts table — requests a change to the validation, and merges after CI passes and a second reviewer approves. The task took the agent twenty minutes and the engineer thirty; it would have taken the engineer half a day. The change went through exactly the review and controls a human's would have, which is the governance rule that makes the agent acceptable in a regulated codebase.
The risks are in what the agent can reach and what the review misses. An agent with production credentials, or one that reads a repository containing secrets, is an incident waiting; sandboxing, scoped credentials and secret scanning are the controls. An agent reading a dependency's documentation or an issue comment is exposed to prompt injection, and a poisoned instruction can become a code change; the review is the defence, and reviewers must know it. Plausible-looking wrong code — a subtle validation gap, a race in a payment path — is caught by tests, review and the security scans that already exist, which is why the agent's output is held to the same process. The economics are being measured: throughput gains are real on well-specified tasks and smaller on ambiguous ones, review load rises, and the developer's job shifts toward specification and review. For AI M&A, coding agents are the application category where distribution (the IDE, the developer relationship) and the model (the labs') compete for the value, and the 2025 deals were priced on growth rates the market had not seen in software.
For a fintech, coding agents are a throughput tool inside an unchanged change-control process, with credentials, sandboxes and review as the guardrails; for an investor, they are the clearest evidence that AI applications can reach revenue fast — and that the labs will compete for it.
Example
A fintech with 120 engineers rolls out coding agents under a policy: agents run in sandboxes with no production access; repositories are scanned for secrets first; every agent pull request is labelled, reviewed by two people, and passes the same CI, security scans and change-control gates as human work; and metrics are tracked. After six months: pull requests per engineer up 45%; median cycle time down 30%; review time per PR up 20%; two agent-produced changes caught in review with security errors a human might also have made; one incident where an agent, following instructions in a third-party library's README, added a telemetry call that review caught. Test coverage rose from 71% to 84% because agents were assigned the test backlog. The cost — $180 per engineer per month — against the throughput is a decision the board makes in a minute; the policy is what made it a safe one.
Common mistakes
- Agents with production credentials or access to secrets. - Agent pull requests reviewed less carefully because "the tests pass". - No policy on what agents may read, run or change. - Measuring productivity by lines of code rather than cycle time and defects. - Reading the vendors' growth as a moat when the labs supply the model and compete for the developer.