We’re getting ClevAgent ready for you. Purchases aren’t open just yet, so feel free to look around!

ClevAgent
Rule framework

Our Rules

The rules ClevAgent uses to understand agent behavior, intervene at the right moment, and explain the result.

Note
These Rules cover agent efficiency and security today: duplicate work, stale context, file and folder organization, oversized files, memory hygiene, unchecked code changes, repeated failures, sensitive information, prompt injection, unsaved work, and agent settings. This page explains what each Rule catches, when ClevAgent steps in, and how Estimated impact is calculated.
OptimizationCut wasted work, keep the agent efficient.
R1duplicate-read

Re-reading files already in context

Agents often read the same file range again even though it is still in the current context. ClevAgent uses three levels of response: it interrupts a large, high-confidence duplicate before the extra read; it gives guidance after the turn when the overlap is useful but not identical; and it quietly records small or uncertain overlaps. A changed file or a different range is never treated as reusable old context.

The nudge

This file range is already available in the current context. Reuse the earlier output, then continue the original task.

How we measure

Estimated impact can include the duplicate read ClevAgent interrupted, the retained-context value of keeping it out of the conversation, and repeated rereads the intervention is expected to prevent. Caps and equivalence checks keep the same content from being counted twice.

Future repeats cannot be observed after they are prevented, so ClevAgent labels the total as an estimate. The model reflects the downstream value of one well-timed intervention while bounding every component.

R2stale-reread

Forgetting its own writes

Sometimes an agent reads back a file it just wrote or edited itself. It already produced that content, so the readback usually means it lost track of its own work. This Rule gives post-turn coaching rather than blocking the read in advance, and it backs off if another process may have changed the file.

The nudge

This Read on {file} follows your own Write. Use the content you just wrote instead of pulling it back again, then continue the original task.

How we measure

Estimated impact includes the repeated rereads the coaching is expected to prevent after the first detected readback. Exact-content and generation checks keep the same edit from being counted twice or folded into Duplicate reads.

The agent already produced the content, so the estimate includes the value of keeping future readbacks out of the conversation while preserving safeguards for external file changes.

R3file-path-advisor

Writing files in the wrong place

New files dropped in random folders get lost between sessions. This Rule watches where the agent saves reusable work and suggests a findable home, based on the project and the agent's workspace, in a simple predictable structure that future sessions and handoffs can locate fast. It stays quiet for ordinary source files, configs, and anything already in the right place.

The nudge

`{filename}` may get buried in `{currentDir}`. Move it exactly to `{agentRoot}/{Project}/{Topic}/{filename}` and keep that filename. After that, continue the original task.

How we measure

Rule 3 is logged outside the dollar-savings total because its value compounds over future sessions. Cleaner paths reduce rediscovery work, improve handoffs, and make later agent context easier to load.

ClevAgent keeps reusable artifacts findable without turning every file write into a warning. It only treats placement as hygiene value, not immediate dollar savings.

R4file-length-advisor

Piling everything into one file

Working files that grow too large slow down every later search, edit, and context load. When ClevAgent can confirm a file has gotten unwieldy, it suggests splitting or trimming it. It respects intent: if you asked for one long file it stays quiet, and it skips formats where line count is not a useful signal, like generated assets, lock files, logs, and tabular data.

The nudge

This file is getting large enough to slow future work. If it has real structure, split or index it; if it is a throwaway artifact, stub, archive, or delete it.

How we measure

Rule 4 does not book savings when the suggestion appears. It only counts after the user sends the cleanup guidance and the event meets the accepted-action criteria.

ClevAgent treats long files as future context risk and separates user-approved cleanup from automatic dollar claims. A warning alone is not counted as savings.

R5memory-md-optimizer

Letting memory files balloon

Agent memory files quietly grow until recall gets noisy. On a set interval this Rule runs a light check for obvious cleanup candidates so memory stays sharp. It never edits memory on its own: it shows you a suggestion, nothing changes unless you send it, and any real rewrite needs a second approval.

The nudge

Run a quick memory hygiene check. Identify only obvious cleanup candidates, do not rewrite memory yet, and ask for approval before any change.

How we measure

Quick checks do not claim savings by themselves. Rule 5 savings require user-approved memory changes and verified shrinkage after cleanup cost is accounted for.

ClevAgent distinguishes a harmless hygiene reminder from a verified memory-compression result. Quick checks are tracked separately from real cleanup work.

R6folder-optimization

Crowded document folders

Document-heavy folders become harder for agents to scan and easier for useful work to lose. This Rule notices measured signs of crowding after a new file is added and offers to help organize the folder. It does not move or delete anything on its own; you decide whether to send the suggestion.

The nudge

This folder is getting crowded with documents. Review what can be grouped, archived, or moved, then continue the original task.

How we measure

R6 is tracked as workspace-hygiene value rather than immediate dollar savings. It records the measured folder signal and only treats a cleanup as accepted after the user sends the suggestion.

Cleaner folders reduce later scanning, duplicate creation, and handoff friction without guessing which files are disposable.

R7unchecked-code-changes

Code changed without a successful check

An agent can finish after changing code without confirming that the updated code works. This Rule waits for the task to settle, then lets you know when no successful test, typecheck, or build was observed after the latest change. It never runs checks or changes code on its own.

The nudge

The agent changed code but hasn't checked whether the updated code works.

How we measure

R7 is tracked as verification-risk reduction rather than immediate dollar savings. It records whether a successful check was observed after the latest code change.

If you click Ask to verify, ClevAgent asks the same agent to run the narrowest relevant check and report the exact result.

R8repeated-failures

Repeating the same failed step

An agent can get stuck repeating the same command or edit without changing its approach. ClevAgent watches for the same failure repeating when no successful progress or new direction happened in between.

The first failure stays quiet. After the second matching command failure, ClevAgent privately asks the agent to reassess. If the same command fails a third time, you can ask it to try a different approach or let it continue. Repeated edit failures are handled privately so routine correction attempts do not interrupt you.

The nudge

Your agent may be stuck. The same step failed 3 times with nothing changed in between.

How we measure

R8 does not claim immediate dollar savings. It tracks a recovery opportunity when repeated failures show that the agent may be stuck.

A successful change or a new direction from you resets the count. ClevAgent only surfaces the customer choice for an unchanged command failure that reaches the third attempt.

SecurityWarn about sensitive information and stop risky actions before they happen.
RAsensitive-infoWarns

Mishandled secrets

ClevAgent watches for the moment your agent mishandles an API key, token, or password: pasting it into a prompt, echoing it back, or reading it from a credentials file. On its own, this Rule warns you and does not block.

Security Mode is optional. When you turn it on, detected secrets are blocked before they leave your machine. The secret never reaches the provider. You can approve to send it anyway, or leave it blocked.

RBprompt-injectionBlocks

Content that tries to hijack the agent

ClevAgent inspects content brought into the agent and blocks it when the content contains instructions that try to override the user's task, steal control, or redirect the agent. The suspicious content is stopped before it can steer the session.

The block works across supported Claude and Codex response paths. ClevAgent shows what was blocked without exposing hidden payloads or secrets.

How the block works
ClevAgent's local gateway sits between your agent and the model. When content looks like a hijack attempt, the gateway refuses that content instead of passing it to the agent. It does not alter or fabricate the message, and it does not read your content for unrelated purposes.
RCprotect-uncommitted-workBlocks

Cleanup that could erase recent work

Some cleanup and reset actions can permanently remove work that has not been saved in version history. ClevAgent pauses only when it can prove files would be lost. Nothing is deleted until you choose.

It shows how many files are affected and their project-relative names, never their contents. You can keep your work or allow the deletion.

RDagent-settings-protectionBlocks

The agent changing its own settings

ClevAgent pauses the agent before it changes files that control its own permissions, connections, or automatic actions. Nothing has changed yet.

It shows which settings files are affected, not their contents. You can keep the current settings or allow that specific change.