Evidence note: This article draws on official project descriptions and PlayAgit’s own aggregate records for 166 pipeline runs from August 28–October 7, 2026, as of October 7. Project descriptions establish what their maintainers document; independent confirmation of their operation and broader claims about agentic delivery remains open. PlayAgit’s figures are operator measurements, not an independent audit. Calculations below are explicitly identified as derived.
The central finding is an incomplete publication breakdown. PlayAgit’s own records for 166 runs from August 28–October 7, 2026 report 1,175 publications, with 344 new posts and 332 updates. Derived from those reported figures, the named categories total 676. The supplied records do not explain how that subtotal relates to the publication total.
That accounting problem gives the delivery frameworks a practical context: a gate needs an inspectable decision, and a delivery count needs a documented definition.
Delivery controls: icebox-cli’s gates and AgentCheck’s reviewers
According to its official process documentation, icebox-cli describes six steps: strategy and intake, specification preparation, implementation and tests, workflow controls, evidence review, and merge and release. Its five gates check scope readiness, specification alignment, test evidence, operational guardrails, and closeout criteria, followed by a production feedback loop. icebox-cli’s process description
The same documentation describes gate evidence through pull requests, issue comments, and sign-offs. That makes the recorded basis for approval a useful evaluation criterion alongside the approval itself. icebox-cli’s delivery controls
AgentCheck’s official documentation describes local review through five focused reviewers covering logic, security, style, guidelines, and product behavior. It says they examine staged changes and produce reports grouped by reviewer. AgentCheck’s review model
These descriptions suggest ways to place explicit gates and code review within agent workflows. Whether those practices characterize agentic software delivery more broadly—and how reliably they work in operation—remains uncorroborated by the supplied evidence.
Keeping work connected across agents and context windows
The projects describe different mechanisms for maintaining continuity:
- Dependency tracking: The agentic-sdlc-url-shortener project describes an orchestration engine combining dependency graphs, gates, governance, and observability. Official project description
- Explicit handoffs: VibrationPlan describes an adaptation of the software development lifecycle for stateless AI agents, using specialized roles, handoffs, and living documentation. VibrationPlan’s framework
- Persistent knowledge: Distillery describes an MCP server for persistent, shared team context across AI coding sessions. Distillery’s project documentation
- Fresh execution context: cc-pipeline documents a workflow encompassing specification, research, planning, building, review, fixes, reflection, and committing, with a fresh context window for each step through the Claude Agent SDK. Its project page also carries a May 2026 deprecation notice, so it should be read as a documented design rather than assumed to be a current adoption recommendation. cc-pipeline’s workflow and status
For developers comparing these approaches, the useful question is what survives a handoff: the requirement, the artifact under review, the evidence considered, and the decision that permits subsequent work. Shared context and workflow sequencing should be evaluated against that concrete requirement.
PlayAgit’s August 28–October 7 run records: discovery and scoring
PlayAgit’s own records for 166 runs from August 28–October 7, 2026, as of October 7, report 31,561 discovered topic candidates across four categories and 223 distinct source hosts.
Separately, PlayAgit’s own records for the same 166 runs and date range report 37,320 scored topics.
The supplied aggregates do not establish that discovery and scoring count the same population on the same basis. They therefore cannot support a straightforward discovery-to-scoring conversion rate. The relationship between the counters remains open until their inclusion rules and treatment of repeated items are documented.
Before production: held, rejected, and sent forward
PlayAgit’s own records for 166 runs from August 28–October 7, 2026, as of October 7, report 37,320 scored topics: 14,538 held before production, 16,144 rejected, and 6,638 sent to production.
A derived arithmetic check confirms that the reported dispositions sum to the scored-topic total:
14,538 + 16,144 + 6,638 = 37,320
This reconciliation establishes agreement among the aggregate counters. It does not establish why individual topics received those decisions or whether the decisions were correct. “Sent to production” also should not be treated as proof that a draft was completed or a post delivered.
Evidence and review: claim checks and draft verdicts
PlayAgit’s own records for 166 runs from August 28–October 7, 2026, as of October 7, report 5,053 research dossiers containing 29,565 claims, of which 1,745 were recorded as independently cross-checked and 6,596 as primary-attested.
Those labels describe different evidence relationships. A primary source can establish what a project announced without independently establishing that its promised behavior holds in practice. The supplied aggregate does not explain the evidence status of every remaining claim.
At the draft stage, PlayAgit’s own records for the same 166 runs from August 28–October 7, 2026 report 1,353 quality-gate verdicts: 1,122 labeled published, 138 sent back for revision, and 93 rejected.
Derived arithmetic reconciles those verdict categories:
1,122 + 138 + 93 = 1,353
The draft ledger’s “published” label still requires a documented connection to a delivery record before it can serve as proof of publication. Claim checks, draft decisions, and delivered posts concern different objects.
Where the publication breakdown stops
PlayAgit’s own publication summary for 166 runs from August 28–October 7, 2026, as of October 7, reports 1,175 published posts, including 344 new posts and 332 updates.
Derived from that summary:
344 + 332 = 676
1,175 − 676 = 499
The derived difference is an unexplained accounting remainder, not evidence of an additional publication category. The supplied records do not establish whether the categories are exhaustive, whether the total and breakdown use matching counting rules, or how individual publication events reconcile with draft verdicts.
This limits the conclusion: the publication summary is not fully explained by its supplied breakdown. It does not establish that content disappeared or that a particular component failed.
What the discrepancy changes
Publication totals, draft verdicts, and claim checks should remain separate until their counting rules and record relationships are documented.
A claim check concerns supporting evidence. A draft verdict concerns an editorial decision. A publication record should establish a delivered result. Combining their totals into a single success rate would require evidence that they describe a linked population with compatible definitions.
The unresolved breakdown also limits comparisons with delivery frameworks. PlayAgit’s aggregates cannot establish that any named framework improved output or review quality; the supplied measurements do not connect those tools to the recorded outcomes.
For related context, PlayAgit’s overview of review roles, delivery gates, and evidence discusses the surrounding design questions.
Developer takeaway: inspect what each gate records
A practical evaluation should ask:
- What is counted? Define whether a counter represents unique items, attempts, decisions, updates, or delivery events.
- What does approval cover? Attach the verdict to the exact artifact version reviewed and retain the evidence considered.
- What happens after revision? Document how a new attempt affects earlier decisions and aggregate totals.
- What proves delivery? Connect the accepted artifact to a destination record and a confirmed result.
- Can the report reconcile? State whether categories are exhaustive and whether they can overlap; expose unexplained differences.
These are decision criteria for evaluating a workflow, not claims that the named projects already satisfy them.
A useful proposed follow-up would trace a completed work item from intake through review to delivery, including its revisions and retained evidence. Such a trace would test a connection that aggregate counts alone cannot establish.
How we apply this in Vision
For Vision, the immediate application is to preserve the distinctions visible in PlayAgit’s own measurements.
PlayAgit’s own records for 166 runs from August 28–October 7, 2026, as of October 7, place 1,745 independently cross-checked claims within 29,565 recorded claims; that supports reporting evidence coverage at the claim level, without extending the label to every draft or publication.
PlayAgit’s own records for those 166 runs from August 28–October 7, 2026 report 1,353 draft verdicts, including 1,122 labeled published; those figures should remain identifiable as quality-gate records.
PlayAgit’s own publication summary for those 166 runs from August 28–October 7, 2026 separately reports 1,175 publications, 344 new posts, and 332 updates; its incomplete breakdown should remain visible wherever the total is presented.
The supported application is a reporting discipline: preserve each measure’s source and meaning, and leave the relationship to delivered results unresolved where the records do not explain it.