![]() |
Library for 2D pencil decomposition and distributed Fast Fourier Transform |

Enterprise use of AI in software delivery has moved through two phases. The first put coding assistants in developers' editors, where they suggested lines and functions while a person stayed in control of every keystroke. The second, now under way, assigns whole tasks to autonomous agents that plan, write, test and propose changes with far less supervision.
Engineering services firms have responded in three ways. Some have built their own agents and delivery platforms. Some license agents from specialist vendors such as Cognition or Replit. Others concentrate on governance and talent, leaving tool choice open. The nine firms below cover all three approaches, and the differences between them matter more than any single headline claim.
Quick picks
Exadel publishes resolution rates for its autonomous agent, Exadel Colleague: 80% of low-complexity tasks and 50% of mid-complexity tasks, according to the company, with 40% of backlog stories automated overall. Every task still ends at a Human Review stage, so autonomy covers the work while approval stays with engineers.
Colleague runs inside each Exadel engagement and takes a ticket from Jira through planning, testing, building and shipping to a pull request in GitHub. The firm has 2,000+ engineers and 100+ enterprise clients, and Colleague can be deployed on client servers, in a private cloud or under client-owned keys. Setup is stated at 2 to 4 hours.
Exadel's release also puts numbers on the economics of autonomy: 100+ human-equivalent hours recovered per project and almost 70% lower AI compute cost. Because Colleague is agent-agnostic, those savings do not depend on a single model provider, and the firm holds Preferred partner status in the Claude Partner Network Services Track.
Key features
Pricing: Not published. A Colleague demo and engineer consultations are available.
Cognizant combines its own platforms with a licensed autonomous engineer. Flowsource spreads generative and agentic AI across the SDLC, Agent Foundry and an Enterprise Agentification framework structure agent rollout, and the Cognition partnership adds Devin, alongside Windsurf, now called Devin Desktop.
Governance is formalized through ISO 42001 certification. The company says 30% of its code is generated with AI today, targets 50%, and maintains a library of 350+ reusable agents backed by 1,500+ agentic specialists.
Key features
Pricing: Not published.
Accenture's route to autonomous coding runs through Replit, in which it invested in 2026 as part of a strategic partnership for AI-driven enterprise development. Its own platform, AI Refinery, focuses on agents for business functions, with a no-code agent builder and industry solutions.
The firm's scale, about 786,000 people across 120+ countries, lets it place autonomous coding inside much larger transformation programs. The in-house element is the agent platform; the coding agent itself comes from a partner.
Key features
Pricing: Not published.
10Pearls approaches autonomy through people and governance rather than a branded agent. It builds agentic AI systems for clients, offers self-healing test automation and places pre-vetted AI developers into client teams.
The firm's 1,400+ experts include specialists in AI security, risk and compliance, and its responsible AI governance services help clients set rules before agents gain more independence. No delivery metrics for autonomous coding are published. The company's AI readiness assessment is a practical entry point for organizations still deciding how much independence to give agents.
Key features
Pricing: Not published.
Founded in 2006 and listed on Nasdaq, Grid Dynamics runs autonomy through its GAIN platform for AI SDLC and an agentic meta-harness that brings Claude and Codex together. Forward Deployed Engineers embed with client teams to operate it day to day.
The firm aimed to train and certify 90% of its engineers on Anthropic and OpenAI platforms, and it displays ISO 27001 certification. A major North American home improvement retailer expanded the AI SDLC platform to 500+ more engineers.
Key features
Pricing: Not published.
Capgemini treats autonomous agents as an enterprise workforce that needs managing. Its agentic control plane is designed to govern and orchestrate that workforce, and Outcome Deployed Engineers deliver within client programs.
With 420,000+ team members, the group spans engineering, operations and R&D. The page reviewed does not name a specific autonomous coding agent, which leaves the choice of tool to be settled during engagement. Capgemini's 2025 revenues of €22.5 billion and presence in 50+ countries make it a fit for programs where autonomous coding is one workstream among many.
Key features
Pricing: Not published.
Globant packages autonomy as AI Pods, units in which agents carry out work under the supervision of human experts. Clients order pods through Glob.AI and pay per output or consumption.
The company claims pods are at least 30% more productive than an engineer-plus-AI approach. Claude-powered pods are available, and Globant holds Preferred Services Partner status in the Claude Partner Network. Platform access currently runs through a waiting list, and clients can choose which model their pods use.
Key features
Pricing: Per output or consumption; exact prices not published.
Thoughtworks built its own autonomous development platform, AI/works, which handles spec-driven development, legacy reverse engineering and code and test generation. Agent/works adds a control plane to govern, observe and cost-manage agents.
The company says AI/works operates with any model and any coding agent. It reports 80% faster mainframe code analysis for a multinational retailer, and Gartner named it a Visionary in 2026. AI/works began through a co-innovation program.
Key features
Pricing: Not published.
GlobalLogic's VelocityAI suite includes an AI-Powered SDLC beside Enterprise AI and Physical AI, with Forward Deployed Engineers central to delivery. The combination suits firms where autonomous software agents must work alongside industrial systems.
The combined Hitachi Digital counts about 35,000 professionals and 880+ active clients. The merger into Hitachi Digital Inc. takes effect on April 1, 2027, and VelocityAI is closely tied to Google Cloud and Gemini.
Key features
Pricing: Not published; offerings listed on Google Cloud Marketplace.
Autonomy works best when its boundaries are explicit. Before the first ticket is assigned, decide which repositories agents may change, which they may only read, and which are off limits entirely. Define what counts as a complete task, including the tests that must pass.
Keep a person accountable for every merge. Even platforms with high resolution rates leave a share of tasks unresolved or partially resolved, and someone needs to catch those before they reach production.
Track three numbers from the start: the share of tickets agents complete, the review time each completed ticket requires and the defect rate after release. Together they show whether autonomy is saving time or moving work downstream.
The order reflects each firm's published information on four points: whether it operates its own autonomous coding agent or platform, whether a human review step is defined, what quantified results or client examples it shares, and what governance credentials it holds. Firms that build and measure their own agents, with a defined review step, ranked higher than firms that license agents or describe autonomy only at the strategy level. All figures are company claims. Scale was not scored, since a firm's headcount says little about how well its agents perform on a given backlog, and pricing was not scored because none of the nine publishes rates.
An autonomous agent completes a task end to end: it plans, edits code, runs tests and returns a result without step-by-step human prompting. People still approve the result.
No. In every model here, engineers plan the work, supervise agents or review their output, and they handle tasks agents cannot resolve.
Exadel, Thoughtworks, Grid Dynamics, Globant and GlobalLogic operate in-house agentic platforms. Cognizant combines its platforms with Devin, while Accenture relies on Replit for coding.
Through review stages, test requirements and governance frameworks. Exadel requires Human Review on every Colleague task, and Cognizant holds ISO 42001 for responsible AI management.
Not directly. Each firm defines and measures tasks differently, so published figures are best treated as indicators to test in a pilot.
A defined set of tickets, agreed review rules, baseline measures and a fixed decision point. Short, focused phases show more than long discovery work, and they give internal engineers time to learn how to write tickets agents can complete.
Exadel leads this group for autonomous delivery under human review, combining published resolution rates with a review stage no task can skip and flexible deployment. Cognizant suits large enterprises that want certified responsible AI and access to Devin, Grid Dynamics fits teams that want embedded engineers running Claude and Codex, and Thoughtworks brings dedicated agent governance tooling for organizations scaling agent use across many teams.