Kauche runs a delivery pipeline in which an AI reviewer approves most pull requests for merge, and the team treats that arrangement as normal operation. The case shows how the merge decision, the point where a sprint normally reasserts control, can move to an automated reviewer. The PICNIC account of that workflow reports that 83 per cent of pull requests there are merged on AI review alone, and describes a separate pipeline of collaborating agents producing hundreds of merges in a fortnight. Once that is the shape of delivery, the two-week box is no longer where work waits. It becomes the place where nobody can say who approved what.
How an agent's working loop differs from autocomplete
A Copilot-style assistant proposes a line and waits. A coding agent reads a ticket, plans an approach, edits across a dozen files, runs the suite, repairs what fails and opens a pull request for review, often with nobody watching in real time. Jason Beres, chief operating officer at Infragistics, wrote in Intelligent CIO that the workflow needs rebuilding from the operating model upward, and the reason is that the two tools sit in different categories of risk.
Duration and memory explain most of the difference. Agent sessions run for minutes or hours, sometimes across days, without a person checking each step. They also keep state between sessions, usually through a configuration file committed to the repository that records architecture and conventions once, so prompts can stay short. Beres makes the sharper observation that this file is becoming the real interface between an engineer and the work being done on their behalf.
What the timebox still does, and what it no longer synchronises
A sprint does several jobs at once. It sets a shared goal, limits how much work is open at once, and gives the team a fixed point at which commitments are re-examined. The first survives agents. The second gets harder. The third becomes optional.
The blog.doubleslash.de analysis of sprints in the AI age puts the case plainly: once a system can deliver continuously, the two-week constraint loses its technical necessity, and the bottleneck shifts from writing code to the organisational model around it. The same piece flags the trap. Drawing on the 2025 DORA report, it warns that AI tends to amplify whatever practice already exists, so a team without small batches and continuous delivery can see instability rise.
None of this abolishes the ceremony. The Augment Code guide to AI sprint planning keeps planning human-led with agents advising: they draft estimates, surface hidden dependencies and forecast capacity, while the team retains authority over the sprint goal and the backlog. The 2020 Scrum Guide already allows outsiders to attend planning to provide advice, so the framework does not need rewriting to admit them. The same guide notes that Jira, Linear, GitHub and Azure DevOps now let work be assigned to agents through issues, boards, work items and assignee fields, so agents become participants in the ceremony.
The debate about the method itself is louder than the practical change. InfoQ's report on whether AI makes the Agile Manifesto obsolete sets a Capgemini executive's claim that agentic delivery is too fast for agile against the reply from others that the constraint has simply moved, from collaborating to build towards deciding what to build and validating that it works. Both readings lead to the same operating question, which is where coordination now happens if not in a fixed cycle.
A daily standup can absorb the same treatment without disappearing. Suhas Bhairav's production pattern for agent-assisted standups has an agent listen to updates, extract decisions, surface blockers, open follow-up tasks and record provenance for each summary, while people keep the decisions, the direction and the exceptions. What changes is the reason the team meets. Progress reporting can be automated. Disagreement about scope, evidence and risk cannot.
Where accountability and access have to sit
Two questions follow for anyone responsible for a team. The first is ownership: every action an agent takes should trace back to a named person who approved that scope, whether the tech lead who wrote the ticket or the engineer who merged the result. The second is reach: least-privilege access scoped to the task, because an agent holding production credentials moves faster than any review process was designed to catch. Neither is a tool setting. Both are decisions about how work gets assigned, and they have to exist before the first autonomous run. Designing them after the first incident costs more.
The same Intelligent CIO report raises a third question that teams discuss less often: how senior engineers get made once agents absorb the entry-level work. Generated boilerplate, routine defect fixes and written documentation are the tasks agents do fastest and cheapest, and they are the tasks that used to give a junior developer a route into the codebase. A team that removes that route and hires only senior people solves a short-term capacity problem while creating a long-term judgment problem. The work an agent absorbs this year does not teach anyone how to make next year's architecture decisions. The workflow has to keep a path for that judgment to develop.
The preconditions that have to hold before a team drops the sprint
Removing a timebox is safe only when the surrounding system can carry the coordination the timebox used to provide. These are the conditions that recur across the sources, and each one fails in a recognisable way when it is missing.
Small batches come first, because a change set an agent produced is reviewable only while it stays small enough for a person to hold in their head; the blog.doubleslash.de piece observes that large agent-generated changes overwhelm review and delivery, and links bigger batches to weaker stability in the DORA work.
Tiered review follows, and the Ship, Show, Ask pattern gives it structure. Ship covers small, low-risk changes that go directly into the main branch, with the pipeline acting as reviewer. Show covers medium changes that are noted and then shown asynchronously to the team, so the team learns without anyone being blocked. Ask covers larger or high-risk changes that go through a conventional pull request review.
A pipeline able to act as that reviewer is the third condition, which means tests at every level, quality gates, end-to-end linting, automated architecture checks and consistent automation, all fast enough to keep pace with the volume an agent generates. A pipeline designed for a handful of commits a day becomes the binding constraint the moment agents are switched on.
Work-in-progress limits replace sprint commitments as the mechanism that stops the queue growing, and a continuously prioritised backlog replaces sprint planning as the mechanism that decides what comes next.
Agent-readable intent is the fifth condition. An issue has to state why the work exists, what success looks like and what must not change, and it must also say what to build. The PICNIC account of harness engineering places that configuration work between business requirements and development guidelines, which makes it a joint responsibility of product and engineering.
A named owner for every agent task is the sixth, with the scope approved before the run starts and recorded where the merge can be traced back to it.
An audit trail is the seventh: logs that tie an output to the input events, the retrieved context and the model version used, plus a tested route back to the last known good state when a run goes wrong.
Which numbers tell you the change is paying off
Tickets closed and lines shipped were weak measures before agents and are worse now, because volume is the cheapest thing an agent can produce. Cycle time, defect rates and real throughput are the measures that survive scrutiny, and they need to be read against whatever productivity story is accompanying the rollout.
The evidence base is mixed enough to keep anyone honest. The Augment Code guide summarises the 2024 DORA findings as improved documentation, code quality and individual productivity, alongside weaker delivery throughput and stability, with the negative results tied to larger batch sizes. The 2025 update found AI adoption correlating positively with throughput while continuing to correlate with higher instability. That combination is the point: speed and instability can arrive together, and a team that measures only the first will misread the second.
The measurement question has a second edge. An agent that can generate volume on demand makes every output metric easier to inflate. Throughput, story points and lines of code all reward the thing agents do best. Cycle time, defect rates and the rate at which changes need rework are harder to game and closer to the experience of the people who have to maintain the result. The Suhas Bhairav pattern suggests tracking sprint predictability, blocker time and backlog aging as well, and tying them to on-time delivery and defect escape rate. Those are operating measures. They tell a team whether the new workflow is reducing waiting and rework, and they give a product owner something to compare across sprints once the sprint itself stops being the container.
The answer to the title question is not a clean yes. The daily sprint survives as a coordination habit in teams that still need a shared goal and a regular check on commitments. It loses its role as the place where work waits, and it loses its authority as the only point at which a merge decision is reviewed. The teams that come out ahead will treat agents as participants in a workflow they have designed, with ownership, access, review tiers, audit trails and a talent path that keeps producing people who can make the calls an agent cannot. That is the operating decision the source reporting keeps returning to, and it is a more useful question than whether the ceremony itself disappears.
The reporting from Intelligent CIO informed the account of coding agents and engineering workflow above.
Sources
- AI delivers in hours. Why do we still plan in sprints?
- AI Sprint Planning With Agents: A Practical 2026 Guide | Augment Code
- Does AI Make the Agile Manifesto Obsolete? - InfoQ
- AI agents in daily standups: practical production | Suhas Bhairav
- The Day Only Requirements Definition Remained: Rethinking the Essence of Scrum in the Age of AI Agents|PICNIC
the workflow needs rebuilding from the operating model upward
Source: From Copilot to Coding Agents: The new engineering workflow CIOs need to design, Intelligent CIO


