BuildNot yet confirmed elsewhere1 publisher2 min readPublished
GitLab's /goal mode lets one model write code while another judges whether the goal is met
GitLab 19.4 adds a /goal command to Duo CLI that loops until a second model judges the objective met or an iteration limit hits. The release adds finer AI spend controls and keeps agent actions inside existing project permissions.
The Engineer · Build desk
Developers can stop a /goal run, change the goal or redirect the agent. A second model judges if the result meets the goal; it is not formal verification. Group owners set default models and restrict them. Administrators can change MCP tool rules. Owners set per-user caps and overrides.
- capability Developer Can stop a /goal run, change the goal or redirect the agent, claim 4
- constraint Second model Judges whether the result meets the natural-language goal; not formal verification, so clear acceptance criteria matter, claim 6
- decision Group owners Set a default model for a feature and restrict the models available to teams, claim 10
- decision Administrators Can change the MCP tool rules at group or project level, claim 13
- capability Owners Set a default per-user cap and individual overrides from a settings page, without the GraphQL API, claim 18
| Who | How | Kind | Claim |
|---|---|---|---|
| Developer | Can stop a /goal run, change the goal or redirect the agent | capability | 4 |
| Second model | Judges whether the result meets the natural-language goal; not formal verification, so clear acceptance criteria matter | constraint | 6 |
| Group owners | Set a default model for a feature and restrict the models available to teams | decision | 10 |
| Administrators | Can change the MCP tool rules at group or project level | decision | 13 |
| Owners | Set a default per-user cap and individual overrides from a settings page, without the GraphQL API | capability | 18 |
What happened
- GitLab Duo Agent Platform also gains three hosted open-weight models: Kimi K3, MiniMax M3 and GLM 5.3.
- In the expanded MCP server tools, read-only operations default to Always Allow, while write and delete operations default to Always Ask.
- Credit usage exports now break out every billable event separately, recording the user, project, session and flow type along with the credits and tokens it consumed.
- Owners can now set a default spending cap per user, plus overrides for individual users, from a new settings page instead of going through the GraphQL API.
Why it matters
- constraint 'Verified' means a second model's judgment of a natural-language goal, so the developer's acceptance criteria set how meaningful a pass is.
- decision Choosing a default model for each feature and restricting teams to a list now falls to group owners, who will need their own pilot data to set it.
- exposure The Always Ask checkpoint on writes and deletes is a default that administrators can change at group or project level, so it protects only while the policy is left alone.
The command is in public beta [2]. A developer can stop a run, change the goal or redirect the agent while it works [4], and the work runs locally with the project's existing context and controls [5]. The output we would watch in the beta is the one where the command reports work it could not resolve [7]. In our view, a loop that can say it is stuck costs less than one that keeps editing until it hits the limit.
ReleaseTEAM, a DevOps consultancy and GitLab partner, cautions that "the practical risk is automation" [14]. Its example is a flow attached to every pipeline failure or new merge request, one that might fire hundreds of times daily [15]. Its advice is to put usage caps in place before wiring a flow to a high-volume trigger, then to track consumption during a small pilot [16].
Developers can see their own consumption, and billing managers get warnings when on-demand use reaches 50% or 80% of a monthly cap [19]. Those thresholds sit on a monthly cap, so they track a trend more than a single run.
GitLab says the three open-weight models give up to four times as many calls per GitLab Credit as many comparable frontier models [9]. At that best case, a call draws a quarter of the credits [21]. The figure transfers to a given workload only if the cheaper model finishes the same task in about the same number of calls. In a loop that is the open question, because an implementer that needs more iterations eats the discount. According to InfoQ, the announcement does not include model-by-model benchmark results or absolute per-call prices [11].
GitLab ties the permission rules and the spend controls together. Manav Khurana of GitLab said: "The platform running the automation is what governs which tools an agent can touch and attributes what it consumes." [20]
What to watch
- Whether GitLab publishes benchmark results or per-call prices for Kimi K3, MiniMax M3 and GLM 5.3, which would test the four-times claim.
- What the default iteration limit for /goal is, and whether administrators can lower it.
- What a per-user cap does to a /goal run already in progress when the cap is reached.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+15
- Incentives60
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
GitLab released version 19.4 of its DevSecOps platform, adding a goal-driven mode to GitLab Duo CLI, three hosted open-weight models, a larger set of MCP tools, and more detailed controls for AI spending.
- [2]
The new /goal command in GitLab Duo CLI, now in public beta, accepts an open-ended objective and continues working until it completes the task or reaches an iteration limit.
- [3]
One model implements the changes while a separate model checks progress against the stated goal.
- [4]
A developer can stop a /goal run, change the goal, or redirect the agent.
- [5]
The /goal work runs locally and retains the project's existing context and controls; the release keeps agent actions within the permissions and approval rules already used by a project.
- [6]
GitLab's use of 'verified' means the second model judges whether the result meets the natural-language goal; it is not formal verification of the code, so clear acceptance criteria remain important.
- [7]
The /goal command can report work that it could not resolve.
- [8]
GitLab Duo Agent Platform gains support for the hosted open-weight models Kimi K3, MiniMax M3 and GLM 5.3.
- [9]
GitLab says the hosted open-weight models provide up to four times as many calls per GitLab Credit as many comparable frontier models.
- [10]
Group owners can set a default model for a feature and restrict the models available to teams.
- [11]
The announcement does not include model-by-model benchmark results or absolute per-call prices, so customers will still need to test the options against their own workloads.
- [12]
In the expanded GitLab MCP server tools, read-only tools default to 'Always Allow' while write and delete operations default to 'Always Ask'.
- [13]
Administrators can change the MCP tool rules at group or project level.
- [14]
ReleaseTEAM, a DevOps consultancy and GitLab partner, cautions that "the practical risk is automation".
- [15]
A flow attached to every pipeline failure or new merge request could run hundreds of times a day.
- [16]
ReleaseTEAM recommends setting usage caps before connecting a flow to a high-volume trigger, then measuring consumption in a limited pilot.
- [17]
Credit usage exports now include a row for each billable event, with the user, project, session, flow type, credits and token counts.
- [18]
A new settings page lets owners set a default per-user cap and add individual overrides without using the GraphQL API.
- [19]
Developers can see their own consumption, while billing managers receive warnings when on-demand use reaches 50% or 80% of a monthly cap.
- [20]
"The platform running the automation is what governs which tools an agent can touch and attributes what it consumes."
- [21]
At the best case GitLab cites, a call on these models draws a quarter of the credits that a call on a comparable frontier model does.
Sources
1 independent publisher whose own reporting we read for this story.
- infoq.comGitLab 19.4 Adds Goal-Driven Agents and Tighter Cost Controls
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.