Establishing what is known
Facts already held, information from your systems, values the runtime can calculate, and details it needs to ask the caller for. Your program defines where each fact can come from.
HOW IT WORKS / INSIDE APOLLO-1
Language understands.
Logic decides.
Neural interpretation and symbolic execution work together in a unified computation. The neural model proposes a reading or reply. Your program determines which facts can be accepted, which actions are permitted, and which outcomes can be reported.
THE ARCHITECTURAL INVERSION
Change what sits
at the center of the decision.
An LLM-first agent puts a model inside a system of controls. Apollo-1 gives executable business logic decision authority, with neural understanding and expression in the same runtime.
01 / LLM-FIRST
Controls around the model.
Controls accept, reject, or retry.
02 / APOLLO-1
Logic at the center.
Language throughout.
Neural understanding provides context.
↳
What changes in practice? If a refund requires confirmation, an LLM-first design relies on the surrounding controls to catch a proposal that skips it. In Apollo-1, the symbolic logic at the center makes confirmation a condition of permitting the refund.
INSIDE THE COMPUTATION
What the program controls.
What the agent accepts as fact, which actions it can take, and what it can say.
Taking action
Rules and formulas execute against established state. Required conditions and confirmations must hold before a write or service call. The returned result establishes what happened. The language model cannot override the declared conditions for action.
Sending a reply
Mechanical checks compare the draft with the record. Depending on the check, the runtime can request a correction, use authored wording, or record a breach. The trace exposes the outcome.
These are points of control within one integrated computation. “One forward pass” describes the architecture; it does not mean one message completes a task or skips tool calls and reply checks.
APOLLO-1 IN ACTION
Declare the logic.
The runtime finds the path.
An employee changes a training reimbursement request. Follow how Apollo-1 establishes eligibility, recalculates the amount and checks authorization before acting. This example uses an illustrative company policy.
THE CONVERSATIONTURN 01
EEmployee
“Can I expense the $2,400 data analytics course?”
AApollo-1
“You’re eligible for learning support. Under your policy, we can reimburse up to $1,500 of this course. Your share would be $900. Manager approval is required before I submit it.”
Employee identity, tenure, annual spend and policy version remain available as the request changes.
THE DECISION RECORDv2.3
ELIGIBLE · ANNUAL CAP APPLIED
- Facts
- HRIS: active employee, 14 months of service. Learning catalog: approved course, $2,400 fee. Benefits ledger: $500 already used this year.
- Policy
- Illustrative company policy: reimburse 80% of approved course fees, up to $2,000 per year. Verified manager approval and employee confirmation are required before submission.
- Action
- Calculate the permitted amount. Do not submit or reserve budget while approval and confirmation are missing.
Grounded in sources
HRIS → employee E-1042; active; tenure 14 months. Policy → minimum tenure 6 months; rate 80%. Policy → annual cap $2,000. Benefits ledger → used $500; reserved $0. Catalog → approved course A; fee $2,400.
Computed explicitly
14 ≥ 6 months → eligible available = $2,000 − $500 − $0 = $1,500 rate-based amount = $2,400 × 80% = $1,920 permitted = min($1,920, $1,500) = $1,500 employee share = $2,400 − $1,500 = $900
Checked before action
employee_eligible = true course_approved = true manager_approval_verified = false employee_confirmed = false submit_permitted = false
Execution events
understood → intent: learning reimbursement learned → employee, tenure, course, allowance sources → HRIS, catalog, ledger reads: found did → report permitted amount submission write → none
THE EXECUTION TRACE
Every turn carries evidence.
Every record names its version.
An LLM can generate an explanation without establishing what caused an answer inside its weights. Apollo-1 exposes the explicit facts and symbolic rules that governed a decision, tied to the program that ran.
RUNTIME v2 / STRUCTURED RECORD
Find the cause.
Make the right change.
A wrong answer can start with interpretation, source data, policy, or a service call. The trace separates those causes so your team can change the relevant behavior.
Explore the trace documentationIdentity
turn · programThe turn, program ID, and exact version that ran. Evidence stays tied to the behavior that produced it.
Understanding
understood · learnedFacts interpreted from the message, values established, and where each came from.
Sources
sources · passages_used · citations_usedReads, writes, and their outcomes. Distinguish retrieved knowledge from the passages actually used or cited.
Decisions
decided · refused_byThe capability, policy, and record evaluated; whether the condition held, its reason, and the policy behind a refusal.
Action & reply
did · checks · answerCapabilities executed, reply-check results, and the answer sent, including any escalation.
Readable account
narrativeA plain-language account alongside the structured execution evidence.
Documented current-runtime fields. Facts marked sensitive are masked.
Evidence across versions.
The build bundle keeps the brief, case bank, and scenarios alongside the program. Test runs supply transcripts, traces, and verdicts. New behavior gets new evidence; earlier results apply to the version tested.
What determines response time?
CPU-based logic and focused language work are part of the picture. Model inference, source lookups, and service calls also matter. Evaluate first-response and task-completion latency on your scenarios.
ENTERPRISE DEPLOYMENT
Deploy on your terms.
Use Apollo-1 as a managed service or arrange a custom enterprise deployment aligned with your infrastructure, security, and data-residency requirements.
Start with the AUI service.
Publish an agent and connect through the API or SDK. AUI hosts the runtime so your team can focus on the program and its integrations.
Align with your environment.
Discuss the deployment and region your organization requires. The agent program and tooling stay the same.
%20--%3e%3csvg%20version='1.1'%20id='Layer_1'%20xmlns='http://www.w3.org/2000/svg'%20xmlns:xlink='http://www.w3.org/1999/xlink'%20x='0px'%20y='0px'%20viewBox='0%200%20304%20182'%20style='enable-background:new%200%200%20304%20182;'%20xml:space='preserve'%3e%3cstyle%20type='text/css'%3e%20.st0{fill:%23252F3E;}%20.st1{fill-rule:evenodd;clip-rule:evenodd;fill:%23FF9900;}%20%3c/style%3e%3cg%3e%3cpath%20class='st0'%20d='M86.4,66.4c0,3.7,0.4,6.7,1.1,8.9c0.8,2.2,1.8,4.6,3.2,7.2c0.5,0.8,0.7,1.6,0.7,2.3c0,1-0.6,2-1.9,3l-6.3,4.2%20c-0.9,0.6-1.8,0.9-2.6,0.9c-1,0-2-0.5-3-1.4C76.2,90,75,88.4,74,86.8c-1-1.7-2-3.6-3.1-5.9c-7.8,9.2-17.6,13.8-29.4,13.8%20c-8.4,0-15.1-2.4-20-7.2c-4.9-4.8-7.4-11.2-7.4-19.2c0-8.5,3-15.4,9.1-20.6c6.1-5.2,14.2-7.8,24.5-7.8c3.4,0,6.9,0.3,10.6,0.8%20c3.7,0.5,7.5,1.3,11.5,2.2v-7.3c0-7.6-1.6-12.9-4.7-16c-3.2-3.1-8.6-4.6-16.3-4.6c-3.5,0-7.1,0.4-10.8,1.3c-3.7,0.9-7.3,2-10.8,3.4%20c-1.6,0.7-2.8,1.1-3.5,1.3c-0.7,0.2-1.2,0.3-1.6,0.3c-1.4,0-2.1-1-2.1-3.1v-4.9c0-1.6,0.2-2.8,0.7-3.5c0.5-0.7,1.4-1.4,2.8-2.1%20c3.5-1.8,7.7-3.3,12.6-4.5c4.9-1.3,10.1-1.9,15.6-1.9c11.9,0,20.6,2.7,26.2,8.1c5.5,5.4,8.3,13.6,8.3,24.6V66.4z%20M45.8,81.6%20c3.3,0,6.7-0.6,10.3-1.8c3.6-1.2,6.8-3.4,9.5-6.4c1.6-1.9,2.8-4,3.4-6.4c0.6-2.4,1-5.3,1-8.7v-4.2c-2.9-0.7-6-1.3-9.2-1.7%20c-3.2-0.4-6.3-0.6-9.4-0.6c-6.7,0-11.6,1.3-14.9,4c-3.3,2.7-4.9,6.5-4.9,11.5c0,4.7,1.2,8.2,3.7,10.6%20C37.7,80.4,41.2,81.6,45.8,81.6z%20M126.1,92.4c-1.8,0-3-0.3-3.8-1c-0.8-0.6-1.5-2-2.1-3.9L96.7,10.2c-0.6-2-0.9-3.3-0.9-4%20c0-1.6,0.8-2.5,2.4-2.5h9.8c1.9,0,3.2,0.3,3.9,1c0.8,0.6,1.4,2,2,3.9l16.8,66.2l15.6-66.2c0.5-2,1.1-3.3,1.9-3.9c0.8-0.6,2.2-1,4-1%20h8c1.9,0,3.2,0.3,4,1c0.8,0.6,1.5,2,1.9,3.9l15.8,67l17.3-67c0.6-2,1.3-3.3,2-3.9c0.8-0.6,2.1-1,3.9-1h9.3c1.6,0,2.5,0.8,2.5,2.5%20c0,0.5-0.1,1-0.2,1.6c-0.1,0.6-0.3,1.4-0.7,2.5l-24.1,77.3c-0.6,2-1.3,3.3-2.1,3.9c-0.8,0.6-2.1,1-3.8,1h-8.6c-1.9,0-3.2-0.3-4-1%20c-0.8-0.7-1.5-2-1.9-4L156,23l-15.4,64.4c-0.5,2-1.1,3.3-1.9,4c-0.8,0.7-2.2,1-4,1H126.1z%20M254.6,95.1c-5.2,0-10.4-0.6-15.4-1.8%20c-5-1.2-8.9-2.5-11.5-4c-1.6-0.9-2.7-1.9-3.1-2.8c-0.4-0.9-0.6-1.9-0.6-2.8v-5.1c0-2.1,0.8-3.1,2.3-3.1c0.6,0,1.2,0.1,1.8,0.3%20c0.6,0.2,1.5,0.6,2.5,1c3.4,1.5,7.1,2.7,11,3.5c4,0.8,7.9,1.2,11.9,1.2c6.3,0,11.2-1.1,14.6-3.3c3.4-2.2,5.2-5.4,5.2-9.5%20c0-2.8-0.9-5.1-2.7-7c-1.8-1.9-5.2-3.6-10.1-5.2L246,52c-7.3-2.3-12.7-5.7-16-10.2c-3.3-4.4-5-9.3-5-14.5c0-4.2,0.9-7.9,2.7-11.1%20c1.8-3.2,4.2-6,7.2-8.2c3-2.3,6.4-4,10.4-5.2c4-1.2,8.2-1.7,12.6-1.7c2.2,0,4.5,0.1,6.7,0.4c2.3,0.3,4.4,0.7,6.5,1.1%20c2,0.5,3.9,1,5.7,1.6c1.8,0.6,3.2,1.2,4.2,1.8c1.4,0.8,2.4,1.6,3,2.5c0.6,0.8,0.9,1.9,0.9,3.3v4.7c0,2.1-0.8,3.2-2.3,3.2%20c-0.8,0-2.1-0.4-3.8-1.2c-5.7-2.6-12.1-3.9-19.2-3.9c-5.7,0-10.2,0.9-13.3,2.8c-3.1,1.9-4.7,4.8-4.7,8.9c0,2.8,1,5.2,3,7.1%20c2,1.9,5.7,3.8,11,5.5l14.2,4.5c7.2,2.3,12.4,5.5,15.5,9.6c3.1,4.1,4.6,8.8,4.6,14c0,4.3-0.9,8.2-2.6,11.6%20c-1.8,3.4-4.2,6.4-7.3,8.8c-3.1,2.5-6.8,4.3-11.1,5.6C264.4,94.4,259.7,95.1,254.6,95.1z'/%3e%3cg%3e%3cpath%20class='st1'%20d='M273.5,143.7c-32.9,24.3-80.7,37.2-121.8,37.2c-57.6,0-109.5-21.3-148.7-56.7c-3.1-2.8-0.3-6.6,3.4-4.4%20c42.4,24.6,94.7,39.5,148.8,39.5c36.5,0,76.6-7.6,113.5-23.2C274.2,133.6,278.9,139.7,273.5,143.7z'/%3e%3cpath%20class='st1'%20d='M287.2,128.1c-4.2-5.4-27.8-2.6-38.5-1.3c-3.2,0.4-3.7-2.4-0.8-4.5c18.8-13.2,49.7-9.4,53.3-5%20c3.6,4.5-1,35.4-18.6,50.2c-2.7,2.3-5.3,1.1-4.1-1.9C282.5,155.7,291.4,133.4,287.2,128.1z'/%3e%3c/g%3e%3c/g%3e%3c/svg%3e){kind=small}
COMMON QUESTIONS
How it works, answered.
How does Apollo-1 enforce a policy?
A policy is a declared condition in the agent program, and the runtime evaluates it before an action is permitted. If a confirmation or an approval is required, the action is not available until that condition holds. The language model cannot override the result, because it never holds the decision.
Will Apollo-1 give the same answer twice?
Given the same established facts, program version and state, the symbolic rules and calculations produce the same outcome. Wording may differ because the language model expresses the result, but the decision, the figures and the actions taken do not.
What is in an execution trace?
Six groups of fields: the turn and program version that ran, what the agent understood and learned, the sources it read or wrote, the decisions it evaluated and any refusal, the capabilities it executed with its reply checks, and a plain-language account alongside the structured record.
How do I change what the agent does?
Edit the agent program, which is typed YAML declaring facts, capabilities, policies, calculations and connections. Describe the change in natural language in the Playground, or edit through the CLI. Re-test against established cases, publish a version, or republish an earlier one to roll back. No domain retraining is involved.
Do I need to learn a symbolic language to build an agent?
No. Describe the task and its policies in plain language and a coding agent can author the program, run scenarios and revise from the trace. Your team still defines the rules, connects the systems and validates the behavior.
Where does Apollo-1 run?
As a managed service hosted by AUI, or as a custom enterprise deployment aligned to your infrastructure, security and data-residency requirements. AUI partners with AWS and Google Cloud. The agent program and tooling are the same either way.
PUT APOLLO-1 TO WORK