"Writing" Agents in Your Editor: Glean Moves Agent Building into Claude Code, Cursor, and Codex

The one-liner: agents finally get a "code-first" way to be built

On October 21, 2026, enterprise AI search and agent platform Glean announced headless agent building as generally available: developers can define, test, evaluate, and publish agents as code — directly inside Claude Code, Cursor, and Codex. An agent becomes a set of files (spec.yaml + instructions.md) that lives in Git, goes through PR review, and rolls back like any other change.

The agents still run on Glean's managed runtime: knowledge graph, permissioned access across 275+ enterprise sources, full audit trail. Glean's pitch: code-first flexibility without rebuilding the enterprise context and governance foundation yourself.

The workflow: four steps, from files to production

Per Glean's official blog and docs, the flow looks like this:

1. Define as files. With the Glean plugin installed, creating an agent in your workspace generates a folder (default .glean/agents/): spec.yaml for configuration (including the agent id), instructions.md for behavior, plus optional skills/ and subagents/. The docs warn explicitly: don't hand-edit the id in spec.yaml — it's the key linking the folder to the live agent in Glean.

2. Test & evaluate locally. Run the agent with representative inputs right in the coding environment, review outputs, run evals, and iterate on instructions and config — the feedback loop stays close to the code. Builders can monitor run progress and inspect output.

3. Review & rollback. Changes can be reviewed in the coding assistant or through Git PRs. A change that doesn't work? Rollback is a git operation — the most tangible dividend of "agents as code."

4. Publish to Glean. When ready, publish to Glean (not a self-built runtime) and share the link with the team. Published agents run on the managed runtime; admins can share, manage, or disable them from the Glean console.

The headless agent building workflow. Summarized from Glean's official blog and documentation.

Product details worth noting:

• Two tracks coexist: users who prefer natural language or the visual low-code canvas keep the agent builder; headless currently supports Auto mode agents only — Workflow mode still needs the visual builder;
• Setup differs per environment: Claude Code via the gleanwork/claude-plugins marketplace (glean@glean-plugins); Codex via terminal commands adding gleanwork/codex-plugins; Cursor via searching "Glean" in Settings → Marketplace; then bind with a work email + Glean OAuth;
• No MCP server for this flow: the docs state the headless builder works only through the plugin's /glean_run chat command ($glean_run in Codex) — don't substitute a Glean MCP server or its tools, and don't reuse beta marketplace/plugin names. Glean is deliberately narrowing the entry point to avoid ecosystem fragmentation.

The official example: birth of a support-escalation agent

Glean walks through a full example: a support engineer wants to catch urgent escalations early. In Codex, they ask Glean to create an agent summarizing support conversations from Slack and Intercom, refine its instructions, test against a real scenario, run an evaluation to find weaknesses — then publish to Glean and share with the team.

The example is well chosen: support-ticket triage is the archetypal "high-frequency, repetitive, needs-enterprise-context" task — exactly the intersection of "code-first building + enterprise governance." A personal knowledge-base Q&A works fine in the visual builder; but an agent reading production systems without permissions and audit is flying blind.

Our take: agent building is undergoing "software-engineering-ization" — vibe coders should steal half of it

On the 2026 timeline, the trend is unmistakable: agents are moving from "prompt craft" to "software engineering." Version control, testing, review, rollback, CI — decades-old software practices are being ported to agent building one by one. Glean's headless is one step: BAND for Developers (which we covered in cycle79) solved multi-agent collaboration; Glean solves build discipline for a single agent.

For vibe coders and indie developers, three takeaways:

First, steal the "agents as files" discipline. You can't use Glean's enterprise runtime, but "keep agent definitions in versioned files, review prompt changes via PR, don't ship without passing evals" applies to any agent project. Next time your support bot misbehaves, ask: are its instructions in Git? Can you roll back? If not, you don't need a better model — you need engineering discipline.

Second, spec.yaml + instructions.md is a file layout worth copying. Config separated from behavior, pluggable skills/subagents — far healthier than "one 2,000-line system prompt." Copy the skeleton for your own agent workflows.

Third, B2B agent builders must learn "governance." Permissions (which enterprise sources the agent can read), audit (a trace of what it did), kill-switch (one-click disable) — these three are the preconditions for enterprises paying for agents. Indie devs selling agent products to companies: the demo can be rough, but these three belong on the roadmap. Glean leading with "275+ source permissions + audit" is no accident.

One cold shower: this is an enterprise product. Its users are "developers at companies with a Glean instance" — indie hackers can't install or use it today. This piece's value is the trend and the methodology, not "go try it." Also note Workflow mode still requires the visual builder — "code-first" doesn't yet cover the most complex orchestration. The engineering-ization of agent building has barely begun.

Transparency

Date sourcing: the Glean blog page shows no publish date; "GA on October 21, 2026" is cross-sourced from unite.ai's Oct 21 report and Glean's docs; the blog text itself confirms "generally available." Per-environment install details come from Glean's official docs (docs.glean.com), cross-checked via unite.ai's summary. 275+ sources, knowledge graph, and audit trail are Glean's official claims. The Workflow mode limitation is explicitly stated by Glean.