Spend a week with a coding agent and you will notice that the hard part is not getting output. The hard part is knowing whether the output works. Agents are fluent, confident and frequently wrong in ways that look fine in a diff and break the moment something touches a real runtime. A run of Show HN posts from the last few weeks suggests the community has stopped treating this as a prompting problem and started treating it as a tooling problem.
The projects below are not models. They are the layer that sits after generation and before you believe anything: tools that drive the running app, grade how an agent experiences your product, review a change by what it does rather than how it reads, and evaluate traces cheaply enough to run on every one of them.
The claim is not the evidence
An agent reporting success is a statement about its own belief, not about your system. The cheapest way to close that gap is to stop asking and start observing. That is the bet behind Reticle, which drives your real running application and reads what actually happened, then returns a verdict of pass, fail or cannot tell with the file and line to fix. The interesting design decision is the third verdict: a tool that admits it does not know is more useful than one that always picks a side, because a cannot-tell result sends you to look instead of letting you ship.
The same question asked from the other direction produces AX Check. Instead of checking your own app, it checks whether an agent can use your product at all: it inspects your site and then has three agents try to onboard end to end, returning concrete guidance on what to fix in your docs, marketing site, CLI, MCP server or Skills files. For developer tool companies whose next customer is an agent rather than a person, that is a category of bug that no human QA pass would ever surface.
Reviewing by behavior instead of by line
Diffs are a poor unit of review. They show you text movement, not consequence, and the larger the change the worse the mismatch. Perspica takes the position that you should review a change by what it does and learn the concepts behind it as you go, which is a better fit for the way agents actually produce code: large, cross-cutting, and internally consistent whether or not the intent was right.
If you want that judgment without giving a tool write access, Jeff is a read-only command line reviewer that uses typed decision models to read a diff and report issues. Read-only matters more than it sounds. A review tool that can edit is a tool you have to supervise, while one that cannot is a tool you can leave running in continuous integration and read when it complains.
Evaluation has to fit inside the loop
Most teams evaluate agents the way they evaluate models: sample some runs, look at them later, adjust the prompt. That rhythm is too slow for an agent that makes dozens of decisions per session. Jevals runs agent evals and guardrails as typed decisions with one request per trace, cheap enough and fast enough to sit inside the loop rather than after it, and it runs locally so traces do not have to leave the machine.
Underneath any evaluation is an environment, and environments are where reinforcement learning runs quietly break. Karotte is a framework for building robust RL environments to train aligned AI, aimed squarely at that layer: not the model, not the algorithm, but the harness the agent learns inside. If your training results are unstable and the model code has not changed, the environment is the first place worth rebuilding.
The harness is the variable you forgot to tune
It is worth separating the model from the system around it. A harness is everything that turns a model into an agent: prompts, tools, memory, skills, orchestration. Harness2 from Google Research is a plug-and-play framework for recursive agent harnessing, built for an open world where agents meet tasks that were never anticipated during harness design, and it ships benchmark guides so the tuning is measured rather than argued about.
Once you accept that the harness is tunable, the next annoyance is that each vendor wants to own it. Endo is the Higher Order Harness, a unified chat interface that composes any agent on the subscription you already have into any workflow, which is a different answer from the usual one: not a better agent, but a neutral place to arrange the agents you already pay for.
Choosing where to start
If your agents ship code nobody verifies, start with runtime evidence rather than more review. A verdict that points at a file and line will change more than another prompt will. If your product is consumed by other people's agents, the audit is the missing piece, because you cannot fix an onboarding path you have never watched an agent attempt. If you are training rather than prompting, the environment layer deserves attention before the model does.
The pattern across all eight is that they refuse to take the agent's word for it. Generation got cheap and confident; verification is where the remaining work is, and it is finally getting tools of its own.
