Harness has rebuilt its software delivery platform so that AI agents can do real work inside it. In a post on 5 November 2026, Rohan Gupta, product lead for Harness AI, set out a new platform experience with a single pipeline model, one Delegate for every capability, versioned step templates and agents that run under the same approvals and policies as people. It is open now through an early access programme, alongside a wave of quarterly product updates published the day before.
What changed
The biggest change is structural. Build, deploy, security, infrastructure and agent steps now share one stage and execution model. Harness has removed the separate stage types for each product, and each step behaves like a reusable plugin. The pipeline YAML is flatter, with sensible defaults. Expressions written in JEXL still work, and CEL is added as a second option. Pipeline inputs can now be lists, structured objects, secrets or connectors, and secret values are resolved with the permissions of the person running the pipeline, so they never sit in the YAML.
The Delegate, the small service that runs work inside a customer's own network, has been rewritten in Go as one agent for all Harness capabilities. Each deployment step now runs in its own container. That means teams no longer install and patch deployment tools on the Delegate host, and step changes ship as new step images without a Delegate upgrade.
A global template library publishes Harness steps and agent steps. Teams can approve a step version, pin it, and use policy (OPA) to restrict which versions are allowed. Different pipelines can pin different versions, and the pinned version is written into the YAML.
The agent becomes the main way to work with the platform. It can write and edit pipelines, explain a failed run, suggest a fix and open a pull request, but each action waits for a person to approve it. It uses the Software Delivery Knowledge Graph, which links commits, builds, services and deployments. Outside coding tools such as Claude Code, Cursor and Codex can connect through the Harness MCP server or the command line, inside existing permissions.
Worker agents are treated as normal resources and pipeline steps. Their actions go through the same approval steps, policy checks and audit log as a human user.
Dashboards move onto the Harness data platform. Existing Looker dashboards are not migrated automatically, and Harness says metric parity is still being checked release by release.
Old and new experiences run side by side. Converting a pipeline is optional and creates a copy, so the original stays as it is. Harness has not given a general availability date.
The quarterly pipeline update adds smaller but useful features: pipeline identities that sign in to OIDC-capable systems without long-lived keys, the option to run policy checks on your own Kubernetes or VMs, and a blast radius test that tries a new policy against recent runs before you save it.
Why it matters
Coding assistants have made writing code faster, but the slow part of most releases is still everything after the pull request: tests, security scans, approvals, change records and rollbacks. If agents are going to help there, they need the same guard rails as people. Harness is trying to put agents inside the controls that platform teams already trust, rather than beside them.
The single Delegate and containerised steps also matter for operations. Many teams run Delegates on hardened hosts in regulated networks, and keeping tool versions current on those hosts is tedious. Moving tools into step containers reduces that work and makes the runtime easier to audit.
Pipeline identities are a quiet security gain. Long-lived cloud keys stored as pipeline secrets are a common finding in audits and a frequent route in supply chain attacks.
There are costs. A new pipeline model means retraining, rewriting some templates and checking that every integration still works. Dashboards that leadership relies on may not match exactly at first.
What this means for organisations in the Middle East, India, Africa and North America
Banks, telecoms operators and government bodies in the Gulf, India and South Africa often run delivery tooling inside their own networks for data residency reasons. The rebuilt Delegate and the option to run policy checks on your own infrastructure fit that pattern, because sensitive checks can stay inside the perimeter.
Regulators are starting to ask who, or what, changed a production system. Central bank technology risk rules in the UAE and Saudi Arabia, RBI guidance in India and OSFI's technology and cyber risk guideline in Canada all expect clear change control. An agent that can only act after a named person approves it, with every step in the audit log, is far easier to defend than an assistant working through a developer's personal token.
In our view, the early access period is the right time to test, not to migrate. Pick one service with a simple pipeline, convert a copy, and see how the agent behaves against your real approval rules.
What to do next
- List where long-lived cloud credentials are stored in your pipelines today and plan a move to pipeline identities.
- Ask your Harness account team for early access and convert a copy of one low-risk pipeline.
- Write OPA policies that limit what worker agents can do in production before you switch them on.
- Check which Looker dashboards you depend on and compare them with the new dashboards before relying on either.
- Decide which outside coding agents may connect through the Harness MCP server, and under which roles.
To plan a pilot, see our Harness partner page. For guard rails around AI agents in delivery pipelines, see our AI security services.