Platform requirements

This page names what a platform must do, never a product. Each need is a capability, and each prompt in this guide lists which of the ten below it needs.

Use the list to judge an agentic platform: software in which an agent can use tools, read and write files and run steps. Kinds include a chat assistant that can call tools, a command-line coding agent (it reads and edits files and runs commands), and a workflow runner that can start sub-sessions. A session is one working conversation with an agent. A sub-session is a new one that a running session starts for part of the work. The reference implementation ran in coding-agent sessions.

The ten capabilities

The table gives the meaning of each capability. The last two columns are suggested options, not tested ones.

Capability	Meaning	Suggested way to supply it	Suggested fallback if it is missing
`llm`	A language model that follows written instructions	Any instruction-following model	No substitute
`long-context`	A model that accepts a very large amount of text in one request	A model with a large context window	Split the text by heading, or write a hand-off document and start a new session
`structured-output`	Output in a fixed format, such as JSON	Name the format, such as JSON (a plain-text format of labeled values), and check the output with a script	Check by hand
`file-read`	Reading files	An agent with file access	Paste the text into the chat
`file-write`	Creating or changing files	An agent with file access	Copy each output into a file
`shell`	Running commands in a terminal	An agent with a terminal	Run scripts yourself; paste results back
`web-search`	Searching the web	A platform search tool	Search by hand; paste links
`web-fetch`	Downloading a web page from a given address	A platform fetch tool	Download by hand
`subagents`	Starting helper agents for parts of a task	A platform that starts subagents	Do the parts one at a time
`human-approval`	Pausing to ask a person to approve a step	Add a pause that waits for a reply	Read each output before going on

Capabilities by stage and workstream

This section sorts the ten capabilities for each stage in the Pipeline overview and for the certification alignment workstream, which is not a stage. Delivery is outside the framework and not assessed here, so it has no entry. The sorting is this guide’s assessment of the project notes; it is not measured. Stage 1 is published; its pages show these capabilities in use on sample prompts and scripts. The words mean:

Required: a person cannot practically stand in for the capability at that scale.
A person can stand in: a person can do the job at that scale.
Not shown: no source shows the capability used.
Marked (inferred): an inference from the project notes, not a stated fact. Every long-context entry is inferred.

Each entry below sorts the ten capabilities under those words.

Stage 1
- Required: llm, structured-output, file-read, file-write, shell, web-search, web-fetch, human-approval
- A person can stand in: long-context (inferred), subagents
- Not shown: none
- Notes: search, screening (checking each source before use) and conversion (changing a file from one format to another) need a model, web access and scripts.
Stage 2
- Required: llm, file-read, file-write, human-approval
- A person can stand in: long-context (inferred), structured-output, shell, subagents
- Not shown: web-search (inferred), web-fetch (inferred)
- Notes: agents fill fixed templates; no source shows web use.
Stage 3
- Required: llm, file-read, file-write, shell, human-approval
- A person can stand in: long-context (inferred), structured-output, web-search, web-fetch, subagents
- Not shown: none
- Notes: the framework specifies automated checks (Layer 1 of the verification layers). In the reference implementation, agents wrote the reports and a script ran the checks. Supply your own checking script, because the reference implementation’s is out of date.
Stage 4
- Required: llm, file-read
- A person can stand in: long-context (inferred), file-write, human-approval
- Not shown: structured-output (inferred), shell, web-search, web-fetch (inferred), subagents (inferred)
- Notes: the tutor is prompts and knowledge files (the course text it answers from), with no code. You load them into a tutor platform, the tool where an AI tutor runs.
Stage 5
- Required: llm, structured-output, file-read, file-write, human-approval
- A person can stand in: long-context (inferred), shell, subagents
- Not shown: web-search (inferred), web-fetch (inferred)
- Notes: agents write test questions in a fixed JSON format.
Certification alignment
- Required: llm, structured-output, file-read, file-write, human-approval
- A person can stand in: long-context (inferred), shell, web-search, web-fetch, subagents (inferred)
- Not shown: none
- Notes: agents built alignment matrices with no code; web use covers researching video resources and checking their links.

Where people replace tools

Some work stays with a person on any platform.

Blocked downloads: a paywall, a login or an access-denied reply stops an automated download, so a person downloads the document by hand. The reference implementation’s rule is no account creation and no scraping (copying pages in bulk with a program).
Running each step by hand: a person can paste one prompt at a time and restart sessions from a hand-off document.
Approvals: a person answers each gate under the human-in-the-loop design; see Human roles, gates and batching.
Expert review: this is Layer 2 of the verification layers, and a person does it. No capability above stands in for it.

Dependencies that are not capabilities

A platform cannot supply these. You arrange them yourself.

Rights to sources: a paywalled or login-gated file needs a lawful copy.
Certification outline: the list of skills the certifying body (the organization that awards the certification) publishes. Certification alignment maps to it.
Reviewers: subject-matter experts (people with deep knowledge of the topic) who can sign off, meaning formally approve.
A file store with history: version control (a tool that records every earlier version of your files) or back-ups, so a lost session costs one batch of work.
Cost: the reference implementation used a cheaper model only for high-volume steps, never for judgment steps.
Credentials: keep the keys and passwords for services you connect to, such as quiz-form, storage and image services, out of repositories (folders under version control).
Tutor platform size limits: check the tutor platform’s caps on file count, file size and instruction length before Stage 4.

Platform profiles

Three profiles show how the capabilities combine. Each is a design expectation, marked untested.

Chat alone, copy and paste: a chat assistant with no tools. It should be able to (untested) draft one Stage 2 chapter from pasted excerpts. You act as file store, shell and approver.
Single agent with files, shell and web: it should be able to (untested) run Stage 1 in sequence, plus Stage 3 checks and certification alignment.
Orchestrated agents with parallel workers: an orchestrator agent hands work to worker agents that run at the same time. It should be able to (untested) add capped parallel waves, meaning groups of workers started together; expert sign-off stays manual. Worker caps are the reference implementation’s parameters, not universal rules; see Three batching practices.

Risks to plan for

Rate limits: a service can refuse or delay your calls once you pass its rate limit. Cap your retries; the reference implementation’s parameters allow at most 2 retries after a temporary failure.
Prompt injection: a fetched page can hide instructions aimed at the AI. Treat fetched text as data, never as instructions.
Context limits: a long source can overflow the context window. Count the words before you load a source, and read by heading.
Unchecked AI output: expect hallucination and invented citations. Check the output and its provenance.
Stale scripts that overwrite results: read a script and back up your files before you run it. By this guide’s convention, a script that writes or deletes files runs as a dry run unless you pass --write.

Readiness checklist (not a validated instrument)

These are questions to think through before you start. They have no score and have not been tested against outcomes.

Does your platform offer an instruction-following model with fixed-format output?
Can it read and write files, and is there version history?
Who runs commands, searches and downloads: the platform or you?
Who answers each approval pause, and who is the expert reviewer?
Do you have lawful access to sources, and a certification outline if you align?
Have you checked rate limits, cost, tutor platform size limits and credential storage?

Next: Stage 1: Knowledge acquisition.