Make Agent Tooling Explain Its Next Safe Move

This article argues for a small product direction in DAL: treat tool results as continuation contracts, so agents can drain useful work without turning ambiguity into authority.

The gap is between execution and continuation

A tool call usually produces a result such as success, failure, or returned data. That is enough for a script, but not always enough for a modular agent system. A coordinator must also decide whether to retry, ask for input, prepare a review item, or stop.

Today, those decisions can remain implicit in prose, memory, or the agent’s next prompt. That makes handoffs fragile. One agent may interpret “booking unavailable” as a reason to search another provider; another may treat it as permission to change dates. Both are plausible, but only one may fit the principal’s intent.

The missing product surface is not a larger taxonomy of tools. It is a standard way for a tool result to say: this is what changed, this is what did not change, and this is the narrowest next transition available.

A continuation contract should be small

A useful result does not need to become a heavyweight workflow engine. DAL could standardize a compact envelope around tool output:

``text result: success | failed | partial | blocked artifact: path or reference to the durable output changed: facts established by this call unverified: facts still requiring evidence next: continue | retry | review | ask | stop scope: the transitions this result may support reason: short explanation for the disposition ``

For example, a travel research tool might return partial, save an itinerary comparison, mark fare availability as unverified, and set next: review before any booking operation. The research remains useful; the result simply refuses to imply a purchase authority it does not possess.

The contract should describe the boundary, not predict the entire plan. A tool should not need to invent five hypothetical future steps. It should identify the next safe decision and the evidence that decision depends on.

Durable artifacts make the contract inspectable

A continuation contract is only trustworthy if another agent or operator can inspect it after the original call ends. That means the result should point to a durable artifact or a durable log entry, not live only in a chat response.

The artifact can be modest: a markdown research note, a JSON result, or a queue item. What matters is that it records the input scope, timestamp, tool identity, and disposition alongside the output. If a later agent receives only the final prose, it may lose the distinction between “found,” “inferred,” and “approved.”

This is where DAL’s file-oriented workflow is an advantage. A handoff can carry the artifact path and continuation metadata together. Reviewers can inspect the exact output, while coordinators can route the item based on next without parsing a paragraph for hidden intent.

The system should also preserve partial results. A blocked tool call should not force agents to start over if it already produced a useful draft. Marking the unresolved boundary explicitly lets the queue drain around the blocker rather than through it.

Review gates belong on transitions, not every tool

The goal is not to make every tool call ask a human for permission. That would add ceremony without adding judgment. Review should attach to consequential transitions: external writes, purchases, publication, destructive changes, or actions whose scope changed since the last check.

A read-only search can normally continue. A generated itinerary can be prepared. A booking submission should become a review item if the result does not contain a matching approval for the exact traveler, dates, provider, and price boundary.

This distinction also improves automation. The coordinator can safely execute internal, reversible steps while stopping at a clearly named external boundary. “Review required” is more useful than a generic failure because it tells the operator what kind of decision is missing.

What DAL should improve next

The product direction can be implemented in stages:

1. 1. Define the result vocabulary. Keep the initial states and next actions intentionally small. Ambiguous statuses create another interpretation problem.
2. 2. Require durable references. Tool results that produce meaningful work should identify the artifact or log entry where that work can be inspected.
3. 3. Make scope explicit. Record what the result permits the next agent to do and what it definitely does not permit.
4. 4. Route dispositions deterministically. continue can return to the queue, ask can create an input request, review can enter the review surface, and stop can close the branch with a reason.
5. 5. Test changed boundaries. Change the destination, recipient, input, or requested effect and verify that the old continuation contract becomes stale instead of silently carrying forward.

These changes are tooling improvements, not a promise that agents will always choose correctly. They reduce the amount of authority hidden inside natural-language interpretation and give operators a stable place to inspect why work stopped or continued.

Conclusion: make the next move legible

DAL does not need to solve trust by assigning every modular agent a permanent score. It needs to make each meaningful tool outcome legible at the point where work continues. A small continuation contract can preserve artifacts, expose uncertainty, and stop external effects without discarding useful progress.

The concrete next step is to add this envelope to one real workflow and exercise its boundaries: successful research, partial output, missing input, changed scope, and review-required execution. If the queue can tell the difference between “continue,” “ask,” and “review” without rereading an opaque transcript, DAL will have improved trust in the place operators actually feel it—the next move.