Can the AI agent actually finish the work?

What counts as finished work?

Finished work has five visible parts: the promised artifact or completed action exists, the written acceptance checks have results, required approvals are recorded, open exceptions are named, and the handoff arrives inside the agreed delivery window. “The agent said complete” is not an acceptance check.

The finish line changes with the hiring mode. A Day Pass is one 24-hour active window with the role and may carry one substantial assignment or several in-scope asks. A Sprint carries a connected project through a named handoff. A Monthly Retainer keeps a continuing function and its working record current. Recurrence is a reason to choose the monthly mode; it is not a condition for finishing one assignment or project.

The work product comes first. Revenue, ranking, savings, filing acceptance, insurance coverage, and legal outcomes depend on facts and decisions outside the delivery. The work-product promise covers the agreed work, acceptance checks, and delivery window.

The same distinction separates useful output from finished work. A summary can help without closing the job. A role-specific AI agent earns the stronger label only when the provider can name the method, finish line, check, approval, repair owner, and retained record.

Which work has a finish line you can inspect?

A finishable work order pairs every promise made before the clock with evidence available at handoff. If the role, input, delivery, check, approval, or failure path is still vague, the buyer is being asked to purchase possibility.

The anatomy of a fidelic agent connects the role's foundation, normal workflows, work products, quality controls, and limits. The public AI agent directory then makes those facts specific to each role.

The acceptance check must survive the run that produced the work. A fresh reviewer should be able to open the file, trace a material fact to an allowed source, reproduce the measurement, see an unresolved exception, or verify that the handoff opens. A self-report from the same AI system is evidence of what it said, not proof that the job passed.

Why can a convincing run still stop early?

Completion and correctness are separate measurements. A system can be accurate when it reaches the last step and still abandon too many runs to own a business job.

What Forge measured

Forge is a research framework for AI systems that call software tools. Its published test used nine synthetic scenarios in two modes, with 50 runs in every scenario-mode cell. That produced 900 trials for each model, serving path, and guardrail condition.

On the pinned Forge v0.5.0 result, Ministral-3-8B-Reasoning-2512-Q4_K_M, an eight-billion-parameter model running through llama-server's native mode, produced 894 correct completed runs out of 900 with the full framework and reached the terminal step in all 900. The same model and serving path without those checks produced 475 correct completed runs and reached the terminal step in 477. Among the runs that did finish, the bare condition was 99.6 percent accurate.

That last number is the warning. Accuracy calculated only after completion can hide abandonment. The surrounding execution system, including response validation, required-step enforcement, and error feedback, was part of the measured result.

What the benchmarks do not establish

Forge does not test FidelicAI, podcasts, small-business records, customer acceptance, or licensed decisions. Its validators are narrow and several look for required strings or state changes. The current Forge repository also carries later, expanded evaluation versions, which is why every benchmark number needs a version, workload, model, serving path, and denominator.

τ-bench, a research benchmark for tool-using customer-service AI, tested a different boundary: 115 simulated retail tasks and 50 simulated airline tasks that required the correct final database state and required information in the message to the user. Each task ran at least three times. Its strongest 2024 function-calling baseline, GPT-4o, passed 61.2 percent of retail tasks and 35.2 percent of airline tasks on the first-trial measure; repeat success fell further. Those historical model results are not estimates for current products. They support a buying habit: run the same realistic acceptance case more than once, inspect the retained state, and check authorization separately.

Neither benchmark proves that a named fidelic agent will finish your work. They show why a model name, one polished demonstration, or accuracy among completed runs cannot carry the claim. Role-specific evidence still has to come from the work product, acceptance check, approval boundary, and failure path.

How does one podcast episode reach handoff?

For one recorded episode, the promised result can include a checked transcript, an approved story, repaired and mixed audio, measured masters, requested show notes, and an editable handoff another producer can open.

SADIE, FidelicAI's AI podcast producer, supplies one current and inspectable example. Her podcast editing work order begins after recording and keeps the show owner's story approval in the middle of the sequence. The public paper-edit sample uses a fictional podcast and is labeled as such. It demonstrates the method, not a customer result.

Each checkpoint answers a different failure. The transcript check catches speaker and wording errors. The paper edit prevents audio cutting before the story is approved. The repair record protects the speaker's meaning. Loudness, true peak, duration, and file-integrity checks make the master inspectable. Opening the editable package tests whether another producer can continue the work.

Requested show notes belong inside the same episode record. SADIE prepares them from the approved episode and checks them against the approved transcript before delivery. That closes the gap between a finished audio file and the complete release handoff the show owner asked for.

For this scope, SADIE replaces human production work: transcript cleanup, paper-edit preparation, dialogue editing, mixing, show-note drafting, and handoff packaging. She does not replace recording, publication, video editing, rights decisions, the show owner's final editorial judgment, or the whole producer role.

What happens if the agreed work misses?

The current FidelicAI guarantee names the failure owner and the remedy. The customer agreement shown before payment is the controlling text.

• Day Pass or Sprint: when agreed work is missing, materially fails its written acceptance checks, or misses its delivery window for a reason within FidelicAI's control, the buyer chooses one no-charge re-performance of the same work or a refund of that engagement fee.
• Monthly Retainer: each full calendar day that a named delivery or response window is late for a reason within FidelicAI's control earns one thirtieth of that month's rate as a service credit, capped at the monthly fee.

The remedy is a commercial response to a defined miss. It is not a completion percentage and does not make the original output correct.

What can still stop the delivery?

Five conditions sit outside that remedy because FidelicAI does not control them:

1. The buyer supplies required access or inputs late, incompletely, or in a form the role cannot safely use.
2. The buyer changes the agreed role, work, acceptance check, or deadline after work begins.
3. An owner or qualified professional does not make a required approval or licensed decision in time.
4. A third-party service is unavailable for reasons outside FidelicAI's reasonable control.
5. The requested action is unlawful, outside the published role, or blocked by the role's approval boundary.

AI output can be wrong or regress. The promise covers the agreed work, written checks, and delivery timing. It does not promise revenue, ranking, savings, a legal outcome, a coverage decision, filing acceptance, or another outside result.

SADIE's boundary is equally concrete. She does not record, publish, edit video, clone a voice, invent a quote, or make the final editorial decision. The show owner approves the story, sensitive cuts, speaker and rights questions, sponsor claims, final master, and publication.

The safest role is often the narrower one. The security and access boundary should grant only the sources and actions the job needs, and an approval-blocked action should remain visibly open.

Which hiring mode fits the work?

Choose by how the work arrives and what must remain current.

• Day Pass: one 24-hour active window with the role. It can carry one substantial assignment or several in-scope asks. The buyer may reprioritize inside the same role and available time. The day ends with completed work and an explicit record of anything still open.
• Sprint: one connected project carried through its named handoff. SADIE's one-episode production sequence is an example.
• Monthly Retainer: a continuing function with the working context intact. It fits a queue, calendar, register, forecast, or production record that must stay current.

The public rate board shows the current modes for every production role. There are no usage credits, seats, activity meters, or surprise overages. Each mode is still bounded by the published role, the available inputs and access, the approval boundary, and the agreed work.

When should a chat product or a person keep the job?

Use a general AI chat when the work is occasional, reversible, and easy for you to inspect. It can be the simpler choice for an outline, a first draft, a quick analysis, or an assignment whose method you want to direct yourself. The AI agent versus chatbot guide compares that ownership line in detail.

Keep a qualified person in charge when the job depends on licensure, physical presence, negotiation, unfamiliar taste, a consequential judgment, or a relationship the customer, guest, employee, or regulator expects from a person. For podcast production, a human producer may also be the better fit when you need recording, video, publishing, promotion, or a creative partner who can make final story decisions with you.

No system should start paid work when the buyer and provider cannot state what will exist, how it will be checked, who must approve it, and what happens when it stops. A broad capability can still be useful. It is not yet a finishable engagement.