Keep an agent’s working plan in files, but put the work that other agents or humans must rely on in a shared, permissioned record. That split is more defensible than choosing either Vercel Board or Markdown for everything. Files are good at preserving execution state; a shared board is intended to make delegation and access visible across principals.
The practical consequence is easy to miss. Once assignment, status and completion evidence leave agent-local files, their history and access rules become properties of the shared system. If that system cannot show who delegated work, who changed its status and who approved the result, a higher volume of agent-generated merges is not an audit trail.
This placement is an editorial recommendation based on documented properties on each side, not a head-to-head test. No independent evaluation of Board exists yet; Vercel’s launch post is the only source, and it does not compare Board against file-based coordination.
Files are execution state, not a failed collaboration model
The file-based approach already serves a distinct technical purpose. A study of code-based agent harnesses describes PLAN.md, Implement.md and status logs holding milestones, acceptance criteria, validation commands and recovery rules. Agents can reload and update those records across context resets and multi-session execution. The file is therefore part of the agent’s ability to resume and verify work, not merely a human-readable TODO list.
Claude Code uses a different file format for coordination, but makes the tradeoff especially clear. An analysis of its multi-agent design reports that each teammate has an inbox JSON file, with assignments and other messages written under a per-inbox lockfile. The paper says this deliberately uses file locking instead of a message broker or distributed coordination service. It sacrifices throughput to obtain two properties: zero-dependency deployment and full debuggability through plain-text JSON at predictable filesystem paths.
That is direct evidence against treating a board as an automatic upgrade. A local inbox can be inspected without operating another coordination service, and a local plan remains available to the execution loop that needs it. Moving those records elsewhere introduces a dependency and can separate the agent’s working state from the mechanism used to debug it.
The weakness is audience. State that is easy for one agent or harness to inspect may still be invisible to everyone else. The community project superpowers-status exists to present work plans, tasks, steps, subagents, progress and ETA from Claude Code session events and repository files. Its description says it follows subagent-driven plans as well as plan-mode plans, Markdown checklists and Claude Code task lists. That is evidence of a real visibility problem, but also evidence that teams can address some of it with a viewing layer rather than replacing their storage model.
The useful distinction is therefore not “files are private, boards are visible.” It is whether the intended audience can inspect the state with the identity, history and context they need.
What Vercel says Board adds
Vercel’s launch post introduces Board as an “experimental agentic primitive” for coordinating work across agent fleets. Vercel says it added grants for different levels of access for humans and agents, along with two-way sync for tools such as Slack.
The same post describes Board as harness-agnostic and names Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra and LangChain as frameworks that can collaborate “as if they were a single harness.” If that behavior matches a team’s requirements, it addresses a limitation of harness-specific inboxes: delegation need not assume that every participant uses the same agent framework.
Vercel also reports an internal adoption result. It says it rebuilt v0’s software factory on Board after using it with security agents, and that the factory “now merges over 50% of PRs”. That is a self-reported vendor figure: it comes with no published methodology, baseline or review outcomes, and nothing shows that the result transfers to another fleet. It does not show that Board caused the merge rate, or that the merged changes received adequate review.
All of these capabilities are vendor launch claims. Vercel states that vercel board “is now available as an experimental primitive,” with agents reading from and writing to it through the CLI or JavaScript SDK. The post does not state pricing or scale limits, and “experimental” describes the launch rather than promising stability in later releases.
A grant is only as useful as its enforcement boundary, and that boundary is where the launch post is at once most specific and least complete.
Which coordination state belongs where?
The following mapping is editorial synthesis. It uses the documented file tradeoffs and Board’s claimed capabilities to make a deployment decision; it is not a Vercel feature matrix or a tested equivalence between Board, files and external trackers.
| Coordination state | Default location | When to share or move it | Reason |
|---|---|---|---|
| Execution plan and recovery state | Agent-local PLAN.md, implementation notes and status logs | Publish a milestone or handoff summary when another principal must review or resume the work | The executing agent needs durable state across context resets; reviewers usually need the current commitment, not every working note |
| Task assignment | Local inbox for a contained, single-harness workflow; shared board for cross-principal delegation | Use a board when humans or agents outside the local harness must see ownership and changes | A lockfile inbox supports local delivery and inspection, while shared delegation requires a common authorization and visibility boundary |
| Working status | Detailed progress in local files; current shared status on a board | Share when blocked work, dependencies or ownership affect someone else | This preserves debugging detail without forcing every observer to reconstruct progress from each agent’s files |
| Detailed execution logs | Agent-local files or the existing execution record | Reference the relevant evidence from the shared completion record when review requires it | Plain-text state is useful for debugging; a coordination board should not automatically become a transcript store |
| Completion evidence | Shared board, or an external tracker when that is the organization’s record of review | Always share when another principal must accept, release or build on the result | A completed checkbox is weaker evidence than the validation result, reviewer decision and location of the underlying artifacts |
| Durable cross-team ownership and release record | External tracker | Use it when responsibility continues beyond the agent session or belongs to an existing organizational workflow | Agent coordination and organizational accountability can have different lifetimes and audiences |
Two qualifications keep this mapping from becoming a synchronization trap.
First, duplicated state needs an owner. If both PLAN.md and a board can change an assignment, the team must decide which field is authoritative and how conflicts are resolved. Vercel’s claim of two-way Slack sync does not establish conflict behavior for Board, much less for synchronization with a local plan.
Second, completion evidence is not the same thing as a status value. For a fleet permitted to merge code, the useful record includes what was validated, which identity supplied the evidence and who was authorized to accept it. That is a review requirement, not a claim that Board currently implements any particular evidence model.
The audit boundary moves with the shared record
Plain-text files make state inspectable. They do not, on that basis alone, provide the shared permission model or review history a multi-team workflow may require. Conversely, putting status on a board does not prove that its history is complete or that its permissions match the sensitivity of the work.
On the permission machinery itself, the launch post is more specific than most announcements. It shows grants scoped per channel, so an agent can be permitted to “read and post there, and nothing else,” with commands to explain what a grant allows and to take a grant away. It describes per-board access rules, sandbox policies that confine an agent to the channels and permissions an operator chooses, and it states that “each board’s events capture who performed each action.” These are vendor claims rather than tested guarantees, but they speak to attribution, differing read and post permissions and revocable access.
Four operational questions remain genuinely open, and they decide whether a board actually improves the audit position:
- What happens to work already underway when a grant is revoked mid-task?
- How long are events retained, and can reviewers reconstruct the full sequence after a session ends?
- Is approving completion a distinct permission from reading status or posting updates?
- Do the access rules and event capture hold for agents running outside Vercel Sandbox?
The launch post does not answer these, and no independent evaluation exists yet that would.
This is also where the v0 figure needs the strongest caveat. A report that more than half of PRs are merged says nothing by itself about authorization, failed validations or reviewer intervention. I would not use that figure to relax a merge gate. For consequential code, the gate should require a reviewable trail regardless of whether coordination happens in Markdown, Board or another service.
Files buy independence, but neither side is automatically safer
The file model’s zero-dependency deployment is a real advantage, yet file-borne instructions can also be an attack surface. SkillGate examines Markdown skill files downloaded from public registries with a single npx skills add command and no real security screening. It reports that a malicious skill file could silently reprogram agent behavior, exfiltrate credentials, inject backdoors or redirect actions to attacker-controlled endpoints.
That finding concerns downloaded skill files, not locally authored TODO files. It does not show that PLAN.md or Claude Code inbox JSON is commonly malicious. Its narrower lesson is that Markdown can carry instructions an agent acts on, so provenance and trust matter when coordination files cross a supply-chain boundary. A board does not automatically remove that risk, and no published evaluation covers Board’s own security posture.
There is also a context-cost question. The String runtime study reports an 87-task benchmark in which on-demand, Markdown-rendered tool interfaces achieved comparable aggregate success across six models, with a reported difference of +1.3 percentage points, while using 33.5% fewer tokens among completed episodes. Its resident interface remained 53 tokens at any catalog size.
That benchmark measures tool-interface rendering, not task tracking and not Vercel Board. It cannot show that a board is cheaper or that file-based coordination consumes a particular number of tokens. It does support a useful design constraint: keep the resident coordination view small, and render detailed evidence when needed rather than assuming every agent must carry the entire shared history in context.
Finally, coordination architecture can affect failure handling, although the available quantitative evidence comes from another domain. A multi-robot study reports that its federated single-agent architecture contained 89% of failures locally, compared with 51% for CFC and 80% for DHMA. It also reported fewer authority conflicts: 1.04 versus 3.79 and 8.11. Those results should not be transferred numerically to coding agents. They are a reason to test what happens when a board and a local plan disagree, two agents claim the same assignment or a shared service becomes unavailable.
When I would pay for a board
I would stay with files when the executing agent needs resumable state, the participants can inspect the same artifacts, and no external principal needs a separate permission or review boundary. In that setting, adding a service creates synchronization and identity work without a clear audit beneficiary.
I would trial a shared board when assignment, current status or completion evidence must be reviewed across agents, harnesses or humans, and the team can define one authority for each shared field. The trial should include a revoked agent, a conflicting update and a completed task whose evidence is challenged. Those cases test the coordination boundary rather than the ease of creating the first ticket.
I would keep an external tracker as the durable record when ownership and approval already live there beyond the agent session. A board can then be evaluated as a coordination layer, but only after its integration and conflict behavior are demonstrated.
Before selecting Vercel Board specifically, verify pricing and operating limits, which the launch post leaves unstated, and test what revocation does to work underway, how long events are retained, whether approving completion is a distinct permission, and how access rules behave outside Vercel Sandbox. Check synchronization-conflict behavior as well, both between Board and connected tools and between a board and a local plan file. Ask for the methodology and review outcomes behind v0’s reported merge rate. Board’s launch claims make it a relevant worked example; they do not remove the need to prove that the resulting trail is sufficient for the code your agents are allowed to change.
Join the discussion
Share a useful perspective or ask a question about this article.