Original Analysis

Why better AI agents need better ways to prove their work

Company-authored reports from Meta, Microsoft and OpenAI share an acceptance problem: knowing when an agent’s work is usable. Their examples show why testing, permissions and runtime checks deserve attention alongside model capability.

Published

The common problem behind different announcements

TideReach assessment

A coding agent can produce a change, a data agent can assemble a configuration, and an enterprise agent can attempt an action. Each creates an acceptance problem: what evidence is enough to let the next step proceed? Recent accounts from Meta, Microsoft and OpenAI approach different parts of that problem. Taken together, they suggest that the usefulness of an agent depends partly on how well its surrounding system can inspect, constrain and verify its work. [1][2][3][4][5][6][7][8]

TideReach assessment

That connection matters because improving the underlying model and improving the workflow are different engineering tasks. A more capable model can make a difficult task feasible. A dependable acceptance process helps an organization decide which of the resulting outputs it can actually use. The reports provide examples of the second task; they do not establish a universal formula for reliable autonomy. [1][2][3][4][5][6][7][8]

Give predictable decisions a predictable home

Meta’s data-pipeline account separates a deterministic routing layer from specialist agents. The router checks required inputs, duplicate work and workflow conditions, while specialists handle domain reasoning. Its Forum engineering report describes a different setting, with agents operating inside a process that includes builds, tests, visual checks and human approval of shipped work. Both accounts put acceptance conditions around the generated work. [2][1][6][5]

TideReach assessment

The useful distinction is between a decision whose rule can be specified and a problem that benefits from open-ended reasoning. A system can enforce a required field or a permission boundary consistently without asking a model to rediscover the rule each time. Our interpretation is that this can reduce the number of consequential decisions left to an agent, while keeping its reasoning available where it adds value. The benefit still depends on whether the rules cover the failures that matter. [2][3][6][8]

Check the behavior that users depend on

OpenAI’s evaluations make the acceptance target concrete. MultiLangBench compares database contents with a reference SQLite implementation. Its separate COBOL migration evaluation requires byte-for-byte output parity across hidden tests. Meta’s Forum account adds another kind of observation: inspecting simulator state and checking the interface as well as compiling the app. These are different ways of asking whether the system produced the intended observable result. [4][1][7][5]

TideReach assessment

For teams building agent workflows, the implication is practical: define the behavior worth preserving before choosing how to test it. Compilation is useful evidence about a program, but a migration may also need to preserve rounding, output formats and business rules. A passing test suite is only as informative as the cases it covers. Better acceptance tests can therefore change which tasks are suitable for delegation, without making every passing result production-ready. [4][7]

Permissions address a different failure mode

Microsoft’s Project Perception account focuses on enterprise-agent actions, combining identity and permissions with contextual signals and runtime supervision. This addresses a different point in the workflow from checking a completed artifact: the organization needs control over what an agent can attempt and when intervention is possible. Microsoft supplies a security approach, rather than an independent benchmark proving its effectiveness across deployments. [3][8]

TideReach assessment

A useful output check cannot compensate for an action that should never have been permitted. Conversely, restricting permissions does not establish that an allowed result is correct. The analytical connection is that acceptance criteria and action boundaries solve complementary problems. Evaluating an agent product therefore calls for questions about both: what can it do, and what evidence establishes that its permitted work was done properly? [3][4][8][7]

Where this could change competition

TideReach assessment

If stronger verification and controls let teams accept larger pieces of agent work with manageable review effort, part of the product value could move into the workflow around the model. Useful differentiators would include clear traces, enforceable permissions, reproducible tests and recovery when something goes wrong. That is a conditional implication of these designs, rather than evidence that model capability or price has stopped mattering. [1][2][3][4][5][6][8][7]

The evidence still needed

Context

These examples come from company-authored accounts and vendor-run evaluations. Meta’s two reports share a publisher; they are not independent confirmations of the same result. The cases cover different tasks and do not offer a common controlled comparison showing how much each design choice improves reliability or total cost. Their reported outcomes should retain those limits. [1][2][3][4][5][6][8][7]

TideReach assessment

The next useful evidence would compare matched workflows while measuring escaped errors, human review effort, intervention and recovery, alongside cost and completion. Independent reproductions would help establish which results travel beyond the original setting. Until then, the strongest lesson is a design question: before giving an agent a larger job, identify how the system will know that the job was done acceptably—and how it will respond when it was not. [1][2][3][4][5][6][8][7]

Sources and further reading

Numbered references connect each passage to its sources and related TideReach coverage.

  1. Engineering at Meta How Meta Built Forum With AI-Native Engineering Back to text ↑
  2. Engineering at Meta The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do Back to text ↑
  3. Microsoft Official Blog Securing the agentic insider Back to text ↑
  4. OpenAI News Advancing Coding in Low-Resource Programming Languages Back to text ↑
  5. TideReach coverage Meta describes verification-led agent engineering behind Forum Back to text ↑
  6. TideReach coverage Meta separates agent routing from reasoning in its data-pipeline architecture Back to text ↑
  7. TideReach coverage OpenAI publishes multilingual coding and behavior-preserving COBOL evaluations Back to text ↑
  8. TideReach coverage Microsoft extends Project Perception to supervise enterprise AI agents Back to text ↑

Explore more articles