Agentagon Intelligence

Agentagon Intelligence adds curated suggestions to audits, evaluation preparation and measured fixes. It is optional: every workflow remains usable when access is missing or the service is unavailable.

Your coding agent prepares privacy-safe summaries and shows each proposed call. By default it asks for your approval before sending anything to Intelligence. Suggestions help guide its work; local evidence and verification still determine the result.

Start with your coding agent

After obtaining access, use Settings in Codex or Settings in Claude Code. Supply the issued service origin and the name of the environment variable containing your key; keep the key itself outside chat.

Once configured, Audit and Optimize can consult Intelligence, including during their internal evaluation preparation stages. You do not need to prepare request files or run lookup commands yourself. Describe what you want investigated or improved in your normal request.

Configure access

Request access through hello@agentagon.ai. In the web app, enter the issued service origin and API key under Settings → Privacy. The key stays in the running app process; only a generated reference is saved, so restart requires entering the key again. Keep the key outside chat and .agentagon/.

Direct CLI configuration

Request access through hello@agentagon.ai. The documented hosted origin is https://brain.agentagon.ai; configure the exact issued origin explicitly. Keep the issued key in an environment variable or credential store. Save its environment-variable name with agentagon _internal setup --scope user --set intelligence.api_key_env AGENTAGON_API_KEY and the service origin with --set intelligence.endpoint https://brain.agentagon.ai. Project-scoped values can override user defaults in the same config file. There is no implicit service address. Do not paste secrets into coding-agent conversations or store them in .agentagon/.

agentagon _internal status reports whether the key and endpoint are configured without revealing their values. Missing configuration leaves audits, evaluation preparation and fixes available locally. agentagon _internal setup --scope user --set intelligence.access_presented true records that optional access was discussed.

Choose approval mode

The default ask mode shows the purpose, destination, workflow and exact redacted JSON payload, then waits for your approval of that request. Declining skips Intelligence and continues the journey locally. A configured key or the coding host's full-access permissions never imply Intelligence consent.

Choose full access explicitly through Settings to allow Intelligence calls without individual approval. Calls and their exact redacted payloads remain visible. Privacy restrictions and workflow limits still apply. You can switch back to ask at any time; project settings override user defaults.

agentagon _internal setup --scope user --set intelligence.mode full_access
agentagon _internal --workspace CHECKOUT setup --scope project --set intelligence.mode ask

Existing installations default to ask unless this setting is explicitly saved. Credentials are never included in previews.

Use during a workflow

After approval, or under explicitly configured full access, the coding agent sends an abstract summary of the application and the question being investigated. It excludes raw code, traces, credentials, customer details, and protected evaluation material. It may consult Intelligence again when later evidence makes a follow-up useful.

Guidance is advisory. It cannot change your permissions, benchmark, constraints, or execution limits. If the service is unavailable, the workflow continues locally.

Request preparation and direct lookup commands

The coding agent prepares separate short UTF-8 files in ignored .agentagon/ storage. For audit, put project purpose, workflow, broad constraints and high-level available evidence in context, and a specific user ask or abstract investigation objective in focus. For evaluation, use context for applicability and constraints and goal for the evaluation question. For a fix, use required focus for the known issue and optional context for applicability. The CLI reads only files explicitly supplied to lookup. Never upload a saved raw audit, evaluation or fix goal: prepare a privacy-safe summary first.

Exclude raw code/traces, personal information, customer identifiers, private URLs, local paths, credentials and protected evaluation material from both files. For evaluation preparation, protected material includes plans, fixtures, cases, ground truth, private inputs and exact acceptance details. For a fix, exclude the frozen evaluator and candidate-specific private material. Generalize proprietary details. Pattern redaction is additional protection and does not guarantee complete removal of sensitive content.

agentagon _internal --workspace CHECKOUT audit lookup AUDIT_ID --context-file CONTEXT_FILE --focus-file FOCUS_FILE --phase initial --limit 5
agentagon _internal --workspace CHECKOUT eval lookup EVALUATION_ID --context-file CONTEXT_FILE --goal-file GOAL_FILE --phase initial --limit 5
agentagon _internal --workspace CHECKOUT fix lookup RUN_ID --context-file CONTEXT_FILE --focus-file FOCUS_FILE --phase initial --limit 5

In ask mode these commands first return status: "approval_required", an opaque approval_id, and the exact redacted request preview without making a network call. The coding host shows that preview and obtains explicit user consent. Repeat the same command and files with --approve APPROVAL_ID after consent, or --decline APPROVAL_ID to skip it. Do not infer consent from starting a journey. Both options are available on all three commands and cannot be combined.

Approval is single use and bound to the owner state, destination, credential reference, phase, payload and refresh choice. Changes require a new preview and approval. A failed or interrupted request consumes its approval, so retries require fresh approval. An approved dispatch sends the saved redacted payload and emits its preview to stderr before sending; the final result remains JSON on stdout. Full access emits the same visible preview while skipping the approval prompt. Successful local cache reads are marked cached: true and send nothing, so they need no approval. A follow-up or --refresh sends a new request and requires new approval in ask mode.

The ID selects the local owner and the client selects the matching service path: /v1/audit, /v1/eval or /v1/fix. The endpoints share a response contract; audit accepts context and/or focus, evaluation accepts context and/or goal, and fix requires focus with optional context. The fix coordinator owns lookups for a run; do not invoke one for each candidate or reviewer. Each workflow proposes at most two logical lookups for its owner: an initial request and, when useful, one distinct --phase follow_up request after later evidence or measurements. Identical completed requests for the same owner, endpoint and phase reuse receipts; changing a supplied field creates a distinct request. --refresh explicitly repeats a completed request. Requests are not automatically retried. See the shared skill reference for preparation and workflow rules.

Client contract

These implementation details are handled by the coding agent and CLI.

API fields, limits, and saved receipts

The request uses a bearer key. Audit and fix use focus; evaluation uses goal:

{
  "context": "Document Q&A using retrieval, an LLM and caching. Preserve answer quality and access isolation.",
  "focus": "Reduce response latency.",
  "limit": 5
}

{
  "context": "Document Q&A using retrieval, an LLM and caching. Preserve answer quality and access isolation.",
  "goal": "Measure whether the agent handles unsupported questions safely.",
  "limit": 5
}

Audit requires context or focus. Fix requires focus and accepts optional context. Evaluation requires context or goal and rejects focus. Any supplied field must be nonblank UTF-8 text; the relevant fields together must be at most 4,000 characters before and after redaction. Omit unused fields. The previous query request field and --query-file option are not accepted. A positive integer limit defaults to five and values above five are capped at five by the client. The wire contract permits zero to five distinct suggestions, each with exactly id, title and suggestion; the current fix and evaluation routes return at most one strongest qualifying recipe or procedure. Audit guidance may return up to five. The response also contains an opaque knowledge_version. An empty successful response means no guidance qualified.

Configure an origin without credentials, a path, query, or fragment. HTTPS is required except for loopback development. The client appends the workflow path, enforces a 30-second total request deadline, declines redirects, bounds response size, and validates responses. Missing keys or endpoints, rejected keys, rate limits, timeouts, service failures and invalid responses do not prevent the owning workflow. Invalid endpoint syntax produces a configuration error. A lookup reports a busy evaluation or fix owner promptly instead of waiting for its active operation. Requests are not automatically retried.

Receipts retain redacted request fields, responses, suggestion IDs, versions, timestamps and status in local .agentagon/ artifacts. Each receipt is attached to its audit, evaluation preparation or fix run; completed requests are cached per owner, endpoint and phase. Reports identify consulted versions and suggestions where that workflow exposes them. Guidance requires local evidence before it can support a finding, cannot define evaluation ground truth by itself, and cannot override permissions, fixed rubrics, frozen evaluators, hard constraints or execution budgets.

Use agentagon _internal eval status EVALUATION_ID or agentagon _internal status --run RUN_ID to inspect the owner-linked receipt index (intelligence on evaluation state and intelligence_receipts on fix status). Audit reports retain their guidance summary; fix reports and the dashboard keep guidance payloads private. Do not hand-edit receipts or canonical workflow state.

Product telemetry is disabled by default and can be explicitly enabled. When enabled, it records lookup outcomes and returned knowledge IDs without context, focus, goal or local references for all three workflows. Cached rereads do not repeat returned-entry events. The knowledge_investigated and knowledge_cited stages remain audit-only because they refer to local audit evidence and saved findings.

Data processing

Agentagon Intelligence uses the redacted request to provide guidance. We do not retain your request data on our servers.