What is a coding agent? What it changes for software buyers

What a coding agent means for your developers

A coding agent takes a plain instruction, such as “build a customer portal with login and invoice history”. It produces working code across several files, runs the tests, fixes what fails and opens a pull request without a person typing each line. The distinction that matters in the research is this. An AI coding assistant auto-completes code that humans mostly write. An AI coding agent mostly writes and submits code itself, from prompts (2).

Instead of completing the current line, the agent plans a sequence of edits, executes them and iterates until the tests pass. It runs inside a developer’s editor or terminal and works a tool loop: read a file, edit it, run a shell command, check git, repeat. Context Mode is an open-source MCP server built to manage that loop. It tracks file edits, git operations, errors and tasks through hooks, and its platform list includes Claude Code, Gemini CLI, VS Code Copilot, JetBrains Copilot, Cursor, Codex CLI, Kiro, OpenCode, KiloCode and Pi Coding Agent (1). The record it keeps of files touched and decisions made is the session memory. That memory lets an agent pick up where it left off across turns (1). Plain-English definitions of agent, context window and the rest sit in our glossary.

The cost of all that tool use lands in the context window. A Playwright snapshot costs 56 KB, twenty GitHub issues cost 59 KB and one access log costs 45 KB (1). After 30 minutes, 40 per cent of your context is gone (1). When the agent compacts the conversation to free space, it can forget which files it was editing and what task was in progress (1). That is why session memory matters as much as the code. Context Mode claims a 98 per cent reduction in tool output by sandboxing it (1).

Why the review queue is the real story

If your firm commissions custom software, pays for maintenance or keeps an integration backlog moving, the pitch is a lower cost per feature delivered. The evidence so far says the saving shows up in the writing rather than the shipping.

The Harvard study behind this draws on aggregated analytics from more than 700,000 employees at over 700 software development firms (2). It covers 300 million work events recorded from 2021 through March 2026 (2). Introducing AI coding agents raised total lines of code by 30 per cent, commits by 20 per cent and pull requests by 23 per cent (2). Issue and epic resolution rates did not move in a statistically significant way (2).

The extra code lands on the review desk. Average review process time, from pull request submitted to merge, grew 49 per cent (2). The share of pull requests with changes requested nearly doubled and comments per pull request rose 35 per cent (2). The share of workers performing code review grew 14 per cent (2). The bottleneck moves from typing to checking.

Adoption is already broad. 95 per cent of firms in the study have implemented AI coding agents (2). By March 2026, 80 per cent of measured firms used some form of AI code review (2). Even so, AI agents produced only 23.3 per cent of review comments and 10.8 per cent of pull requests (2). People still do the bulk of the checking.

For a UK owner the practical reading is blunt. You may get more code faster, and you will not get faster delivery unless review, testing and deployment change with it. That is why the platform vendors moved. UiPath now lets any coding agent build, test, deploy and govern automations under one orchestration layer. Policy enforcement, audit trails, credential vaults and role-based access control are standard for every automation entering the platform (4). OutSystems opened its platform to Claude Code, Cursor, Codex and Kiro with security, integrated lifecycle management and architectural guardrails (5). OutSystems also cites developer survey data showing 74 per cent of AI-generated code requires significant post-generation rework or is never deployed to production (5). We do this kind of governed automation work for UK businesses.

The results they report are specific. Normal, a Danish retailer, had a single developer build an application in two weeks against a previous month and a half, using Claude alongside OutSystems guardrails (5). YESCO compressed its software release cycles to two weeks (5). Lowenstein Sandler, a US law firm, built a production-ready Litigation Matter Tracker in under three hours. It cut finishing times on that workflow from one hour to five to ten minutes (5).

How a coding agent works, one change at a time

The agent receives a prompt and builds a plan: which files to create or change, what tests to add, what dependencies to install. Then it loops. Propose an edit, run the tool, read the result, decide the next step, until the acceptance criteria are met or it hits a limit.

How it moves

Where a coding agent speeds up, and where it slows

One change request from the backlog, from prompt to merged code.

Comes inA change request: add a new carrier integration

Plan
Which files to create or change, which tests to add, which dependencies to install.
After 01Plan: three files, two new tests
Edit, run, read, decide
It loops: propose an edit, run the tool, read the result, choose the next step.
After 02Loop ends when the tests pass
Judge its own output
Tests, and research shows a second agent reviewing the patch catches more bugs.
After 03Second agent's review: two issues fixed
Pull request
The platform enforces your architectural rules and routes it into the existing review gate.
After 04Pull request opened
Human review
Review time grows by about half as volume rises. The constraint moves here.
A person decides here
Leaves asReviewed and merged

Goes outA change shipped, if review keeps pace

Gilt ring: a person makes the call at this step

The evidence in this guide: lines of code up 30 per cent and pull requests up 23 per cent, while average review time grows 49 per cent. The example change is illustrative.

Session continuity is where most setups fall over. Context Mode persists session state and enforces routing across 17 platforms via MCP and hooks (1). Full session continuity, capture, snapshot and restore, works on Claude Code, Gemini CLI, VS Code Copilot, JetBrains Copilot, OpenCode and KiloCode (1). GitHub Copilot CLI uses its own camelCase hook configuration keys and captures prompt, tool, compaction, session-start and stop events (1). Pi Coding Agent and OMP register tool call, tool result, session start and pre-compaction handlers through extension and plugin hooks (1). Codex CLI gets partial session tracking through hooks, with MCP tools working (1). Cursor captures tool events but its sessionStart hook is rejected by the platform validator, so session restore after compaction is unavailable there (1). Antigravity IDE and Zed have no hook support in the current release, so they rely on manually copied instruction files at roughly 60 per cent compliance (1).

The agent also needs a way to judge its own output, and this is where the second wave of research sits. Meta’s RADAR review tool has handled over 535,000 diffs, of which more than 331,000 were landed (3). It cut median review wall time by 35 per cent, and production incidents fell to one-fiftieth of what manual reviews produced (3). Pairing two coding agents to review each other’s patches improved bug detection more than giving one agent a bigger budget (3). Adversarial Review, tested on the LiveCodeBench benchmark, reached an 87 per cent pass rate, beating single-agent setups and even some configurations using five agents (3). Meta’s just-in-time testing framework plants deliberate faults and checks whether the tests notice. It produced a fourfold increase in bug detection across more than 22,000 generated tests, and up to a twentyfold improvement for significant failures (3). A system called Wink recovered from approximately 90 per cent of coding agent misbehaviours across more than 10,000 real-world instances (3).

Deeper optimisation work runs the same loop at scale. Google DeepMind’s AlphaEvolve pairs Gemini Flash for breadth of ideas with Gemini Pro for depth. Automated evaluators score each proposed program, and it evolves whole codebases rather than single functions (8). In Google’s data centres its scheduling heuristic has been in production for over a year and recovers, on average, 0.7 per cent of Google’s worldwide compute resources (8). It sped up a matrix multiplication kernel in Gemini’s architecture by 23 per cent, giving a 1 per cent reduction in Gemini’s training time (8). It reached up to a 32.5 per cent speedup on a FlashAttention kernel (8). On the mathematics side it found a way to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving on Strassen’s 1969 algorithm (8). Kernel optimisation work that once took weeks of expert effort now takes days of automated experiments (8).

A worked example: forty change requests at a Midlands distributor

Every figure below is illustrative. This is an invented UK business written for this guide, and the numbers are our estimates rather than measured outcomes.

The firm turns over £12 million and runs a custom order-management system built five years ago. The backlog holds 40 change requests: new carrier integrations, pricing rules, warehouse pick-list tweaks. Two developers and one contractor clear perhaps six items a month, and maintenance and bug fixes consume 60 per cent of capacity.

Our estimate is that bringing in a coding agent under a governed orchestration layer lets the team prototype each change in hours rather than days. Two examples of such a layer are OutSystems Agent Experience and UiPath for Coding Agents. The agent writes the initial implementation; the platform enforces architectural rules, runs automated tests and routes the pull request into the existing review gate. Review time per item probably rises at first, but the volume of items reaching review grows faster. If the team ships 12 items a month instead of six, cost per delivered feature falls even with longer reviews. The contractor moves to higher-value integration work.

The downside: the first three months feel slower while reviewers get used to machine-written changes. Budget for platform licences, setup and three months of reduced velocity before the new rhythm settles, and get real quotes, because the sources here publish no prices. Before committing, calculate what automating the task is worth and check the UK cost bands for this kind of build.

What it costs, and what breaks first

The sources here publish no list prices for coding agent licences or orchestration platforms, and we will not invent any. UiPath for Coding Agents is available now to enterprise customers, with initial support for Claude Code and OpenAI Codex and further integrations planned for 2026 (4). OutSystems Agent Experience went generally available on 8 October 2026 (5). For budgeting, start from what custom software costs in the UK, what business process automation costs, then check where AI spend leaks across a team.

The real cost is review capacity. The Harvard data shows average review time growing 49 per cent and the share of pull requests with changes requested nearly doubling (2). Add no review capacity and automate no review, and the backlog simply moves downstream.

The quality gap is documented. AgentSWE is a benchmark of 25 tasks spanning creation, editing and optimisation of agent software (6). It found that no coding agent comes close to delivering what a requirement asks for. The gap between what a builder demonstrates and what it delivers is the norm, not the corner case (6). Recurring failure modes included overfitting to visible test cases, hollow runtimes behind code that looks complete, and unsupported claims of completion (6). SWE-Critix is a pipeline for training verifiers that judge coding-agent trajectories. It improved classification accuracy from 57 per cent for the base Qwen3-30B-A3B reasoning model to 66 per cent after supervised fine-tuning and 78 per cent after reinforcement learning (7). Useful, and still not a guarantee: a verifier’s verdict is another opinion to weigh, not a replacement for your own review.

Governance and compliance: agent-generated code must clear the same static analysis, dependency scanning and penetration testing as human-written code. UiPath’s platform supplies policy enforcement, audit trails, credential vaults, role-based access control and runtime controls for every automation (4). OutSystems adds automated testing and governance guardrails out of the box. It can deploy to public cloud, private cloud, on-premise or hybrid, so data residency gets settled before you build (5). Ask any vendor where the orchestration layer runs and whether your data leaves the UK.

Staff morale: developers who feel replaced will resist. Frame the agent as the thing that handles boilerplate and test scaffolding so your people spend their time on domain logic and architecture. The research points the same way. The Harvard authors cannot attribute significant employment changes to AI, while the share of workers performing code review rose (2).

What to ask the firm that builds your software

If you pay a supplier or an in-house team for custom software, the useful questions are about review, not typing speed.

Do you use coding agents on our work, and which parts of the system do they touch?
Who reviews what the agent writes, and how many hours of review capacity are in the plan?
Which tests run before anything reaches our live systems, and who sees the results?
Where does our code go when an agent works on it, and does it leave the UK?
If the work gets cheaper to write, how does that show up in the price?

Faster writing only pays you back if review keeps pace. A supplier with no answer to the second question has not planned for the bottleneck the evidence points to.

What to try next, before you buy seats

Start with one low-risk, high-volume task: a recurring integration, a report generator, a form-to-database flow. Audit the process before you automate it and map it end to end. Then measure baseline cycle time, review time and defect rate for a few weeks, so you have something to compare against.

Run the agent in a sandbox with the orchestration layer’s governance switched on, and compare cycle time, review time and defect rate against that baseline. If the numbers hold across a full month, expand to a second task. Build an AI roadmap with published UK cost bands. Start from where to begin with AI, then take the five automations UK businesses should test first.

AGMM is an AI consultancy, and custom software is one of the departments of the business we work in; the full list is on our services page. The first step is the AI Constraint Audit: a fixed £2,000, 14-day audit that finds constraints, weak points and opportunities to save or earn serious money before you automate anything. A more extensive audit always follows and is included in the proposal.

Frequently asked questions

What is a coding agent?

A coding agent is an AI tool that plans, writes, tests and refactors code across several files from a natural-language prompt (2). It acts on the codebase itself, using file edits, shell commands and git rather than suggesting the next line (1). An assistant, by contrast, auto-completes code primarily authored by humans (2).

What is a coding agent in GitHub Copilot?

In the material here, GitHub Copilot shows up in two shapes. VS Code Copilot and JetBrains Copilot have full session continuity under Context Mode, covering capture, snapshot and restore, while GitHub Copilot CLI uses its own camelCase hook configuration keys and captures prompt, tool, compaction, session-start and stop events when the plugin hooks are installed (1).

What is an agentic coding tool?

It is the same thing under a different label. UiPath and OutSystems both use coding agent for tools such as Claude Code, Codex, Cursor and Kiro whose output then passes through an orchestration layer with governance and audit controls (4) (5).

What is agentic coding with Claude?

Claude Code gets full session continuity under Context Mode, meaning capture, snapshot and restore across a session (1). It is also the first supported agent in UiPath’s coding agent integration alongside OpenAI Codex (4), and one of the agents OutSystems opened its platform to when Agent Experience reached general availability (5).

How much does a coding agent cost?

The sources here publish no prices, so there is no reliable number to give. UiPath for Coding Agents is available now to enterprise customers (4), and OutSystems Agent Experience is generally available (5). UK price bands for the surrounding work live in what it costs.

Can a coding agent replace a developer?

No. The Harvard researchers cannot attribute significant employment changes to AI after looking at total active workers and cross-referencing LinkedIn data (2). What changed is the mix, with a 14 per cent increase in the share of workers performing code reviews (2). Over a three-month trial at Meta, 80 per cent of the Engineering Agent’s fixes received human review and approximately 25.5 per cent of those reviewed fixes were accepted into production (3). The work moves from writing fixes to judging which ones deserve to ship (3).

If this was useful, these go further on the same ground:

Sources

Every numbered citation in this guide links to one of these. Each was fetched and read on 12 October 2026.

How this guide was made: AGMM’s research engine gathered and read the sources above, a first model drafted the guide from them, a second model checked every sourced sentence against them, and an automatic rule check rejected any unsupported figure or broken link. It is marked for review by a named person at AGMM before it is treated as final. How we write and check our guides.

Answers question 01
Answers question 02
Answers question 03
Answers question 04
Answers question 05