Amdahl Platform API

Resume Agent Session

Resume an agent run that is paused on pending user input. Use when replying to an outline approval prompt, a clarification question, or a continue/finish decision. The input payload must match the run's `pending_input_schema`. After a successful resume the run transitions back to running and proceeds from where it paused.

Cancel Agent Session

Cancel an in-flight agent run. Use when the user wants to stop a running agent before it finishes, or abandon a run that has gone sideways. Cascades to any sub-agents spawned from this run via agents.start — every live descendant is cancelled atomically. The runner bails at the next iteration boundary after observing the status change. Already-completed runs cannot be canceled. Pass a reason to record WHY the run was aborted in the audit trail.

List Workspace Agents

List the workspace agent roster - Amdahl-shipped library agents merged with tenant-authored ones - as lean rows of slug, name, description, and provenance. Read when choosing which named agent should handle a task or rendering the agent directory. Library entries are code-defined and locked; tenant rows carry their creation time.

Create Workspace Agent

Save a new tenant-authored agent: a named, reusable prompt the workspace can dispatch later by its stable kebab-case slug. Use when a user wants to capture a custom instruction set as a workspace agent. Rejects a slug that shadows an Amdahl library agent (those ship in code and cannot be replaced) and surfaces tenant slug collisions as a typed conflict.

Get Agent By Id Or Slug

Fetch one named agent by slug or id, including the full prompt text it runs with. Resolves the Amdahl code library first, then falls back to the tenant row. Read when inspecting exactly what instructions an agent carries before dispatching, editing, or duplicating it. Returns null for another tenant or a missing id.

Update Workspace Agent

Revise a tenant-authored agent's name, description, or prompt in place, keeping its slug stable. Call when refining how a saved agent behaves after seeing its output. Amdahl library agents are code-defined and immutable, so edits targeting one are refused with a locked error rather than silently ignored.

Archive Workspace Agent

Retire a tenant-authored agent by soft-archiving it out of the roster; the row is retained and nothing is destroyed. Use when the workspace no longer dispatches a saved agent. Archived agents disappear from listings immediately, while Amdahl library agents cannot be archived at all - they are code-defined and always available.

List Routines

List the workspace Routines, newest first: each entry carries the prompt, cadence (cron + timezone), agent pin, per-fire config, and last/next fire bookkeeping. Read when rendering the Routines page, checking what refreshes on a schedule, or picking one to update or fire now. Filter by enabled; paginate via limit + offset.

Create Routine

Create a Routine: a cron that fires a Chat each occurrence, running one Master agent turn in a fresh named Chat. Use when the user asks for a standing refresh or a recurring investigation. Provide name, prompt, and cron (plus optional timezone and per-fire config like write_outputs or an agents roster); the fire never pauses on a human and each occurrence opens a new Chat.

Get Routine

Fetch one Routine in full: prompt, cadence, agent roster, per-fire config, enabled state, and the last fire pointers (Chat + run ids). Read when opening a routine detail view or before patching one. Unknown and cross-workspace ids return null.

Update Routine

Patch an existing Routine: rename it, change its prompt or cadence, re-pin the agent, adjust per-fire config, or enable/disable it. Use when tuning a standing scheduled ask without recreating it. Only the supplied fields change; a cadence change is re-validated first and takes effect on the next scheduler tick.

Delete Routine

Permanently remove a Routine so its cadence stops firing. Use when the user no longer wants the standing scheduled ask. Deletion affects only future fires - Chats and documents earlier fires produced are untouched. Prefer disabling (routines.update enabled=false) when the user may want it back.

Run Routine Now

Fire one Routine immediately, off-cadence: opens a fresh Chat and starts the Master turn exactly as the cron would. Use when the user wants the standing refresh right now instead of waiting for the next occurrence. Returns the Chat and run handles (stream/read/resume URLs) so the fire can be watched like any Chat.

Chat

Ask the workspace anything: opens or continues a named Chat and enqueues one Master agent turn. Use when starting an investigation over company data, memory, and connected sources. Always returns handles immediately (chat_id, run_id, stream/read/resume URLs) - subscribe to stream_url or poll read_url with wait_ms for the answer; never blocks.

List Chats

List Chats for this workspace, newest activity first. Read when picking a Chat to continue, rendering a recents rail, or checking what the team has been asking. Each entry is a lean snapshot (name, status, turn count, last activity); filter by status or paginate via limit + offset.

Get Chat

Fetch one Chat snapshot: name, status, its runs (with read handles), which runs are still in flight, and any pending human asks with their resume URLs. Read when rendering a Chat view, deciding whether to continue it, or surfacing an awaiting question to the human. Returns null for unknown or cross-workspace ids.

Rename Chat

Rename one Chat so the thread is findable later. Use when a chat was auto-titled from its first ask and deserves a human name, or when its focus shifted. Renames the underlying conversation title in place; runs, substrate, and handles are untouched. Returns the updated chat snapshot.

Get Chat Run

Read one Chat run snapshot: status, pause payload when awaiting input, the answer envelope (answer_text plus the ordered content_blocks with query and snapshot data), answered asks, and usage. Read when polling a started Chat for its result; pass wait_ms (max 30000) to long-poll until the run settles instead of tight-looping. Pass include=events to also get the persisted activity trace for transcript replay.

Get Chat Preferences

Fetch the calling user's saved Chat settings for this workspace: their default run-config knobs (depth, on_question, write_outputs/memory, external_search, include_divergence, actions_allowed) and their agent block list. Read when opening the Chat settings surface to seed its controls. Returns defaults when nothing is saved yet.

Update Chat Preferences

Save the calling user's Chat settings for this workspace: their default run-config knobs and/or their agent block list (the agents the Master may not delegate to). Use when the user changes a control in the Chat settings surface. A top-level PATCH — only the keys you send change; the block list applies to new Chat turns server-side.

Fast Search

Run a fast, synchronous search in ONE blocking call — the low-latency lane beside Chat. Use when you want an answer NOW rather than a full agent run: it plans the ask into one or more sub-questions, writes SQL over the interactions warehouse for each, and runs them in parallel; in blended mode it also does a quick web + news citation pass. Returns the per-question groups (rows + the SQL that ran) and any citations. Point open-ended, multi-step asks at Chat instead.

Routed Warehouse Search

One routed door over tenant warehouse data: each ask lands on typed structured filters, an NL-to-SQL pass, or meaning-based vector similarity. Use for lookups where you hold a plain question OR typed field predicates and want the cheapest lane chosen automatically (force one via mode). Replies with which lane ran, rows or similarity matches, any compiled statement, and mirror freshness. Discover filterable fields via search_field://list first.

Search Filter Fields

Read when composing typed predicates for the routed search verb: the per-surface field catalog (interactions, deals, deal_qualification) listing each field with its type, description, and the comparison operators it admits. Derived live from the warehouse schema catalog - exactly what the filter compiler validates against, so it can never drift from the accepted vocabulary.

Tiered Company Enrichment

Tiered account intelligence for one company, keyed by domain or by name (a bare name is matched against companies in your own conversation data). Use when preparing for an account: a cached brief returns instantly; on a miss you get the tenant’s own conversation evidence about that account in a bounded read while a background job rebuilds the full brief. Pass mode=full to wait for the complete web + LinkedIn + Crunchbase fusion instead.

Tiered Person Enrichment

Tiered background on one person, keyed by linkedin_url or email (a bare name is refused as unresolvable). Call before an outreach or a meeting: a cache hit returns the stored person brief immediately; otherwise first-party mentions of them come back within seconds and the deep profile refresh is queued behind the scenes. mode=full blocks for the fused profile.

Tiered Topic Enrichment

Tiered market read on a topic or category term. Call when scoping a theme: a fresh topic brief from the store returns at once; on a miss the customer-voice evidence for that topic arrives in a bounded read while the full market fan-out regenerates behind the scenes. mode=full runs the whole divergence-map synthesis inline before answering.

Entity Lookalikes

Find the entities most similar to a seed company or deal, ranked by centroid cosine similarity over the tenant’s own conversation corpus. Use for "more accounts like this one" prospect-expansion asks: give an entity_id (or a domain for companies) and get the nearest neighbours with scores. Honest degradation: when centroids are not materialized yet it says so instead of guessing.

Similar Themes

Rank the tenant’s ML conversation themes nearest in meaning to a free-text description. Use to locate which existing themes an idea, campaign angle, or complaint maps onto — a millisecond centroid match when the fast vector mirror is synced, transparently falling back to the warehouse theme index otherwise. Returns theme ids, labels, member counts, and similarity scores.

Run Eval

Grade content against your workspace data and get a run id to poll. Use when scoring a prompt or a drafted message: pass inputs (the prompt, the message, or both) to the prompt-and-message-eval — it retrieves real customer quotes, an LLM judge scores grounding/specificity/tone with per-dimension reasoning + cited quotes + a rewrite. Returns the run handle (eval_run://<id>) immediately; reuse=cached reuses the last run for the same inputs. Read the scorecard from the run resource.

List Evals

List every eval available to this workspace (prompt-and-message-eval, plus any you authored). Use when choosing which grader to run or rendering an eval picker. Returns slug, name, version, description, input_schema, and case count per eval; fetch eval://<slug> for the full cases + graders.

Create Eval

Create a custom eval — a configured grading pipeline (declared input fields + cases + graders) stored for your workspace. Use when the built-in graders don't fit and you want your own scored report card. Pass slug, name, input_schema, and cases; the whole definition is validated and you get back the new eval's slug + id, or a structured validation_failed with the exact field errors. Built-in slugs are reserved.

Get Eval

Fetch one eval by slug or id, with its input_schema and full cases — each case subject (a provided input field or a generated target) and the graders (rule checks, system-of-record anchors, judge and evidence-judge rubrics) applied to it. Use when building a run form or inspecting what a grader measures. Returns null for an unknown id.

Update Eval

Revise one of your workspace's authored evals in place. Use when tuning a custom grader's rubric, thresholds, input fields, or cases. Pass the slug plus the fields to change; the merged definition is re-validated before it is stored, returning the slug + version or a structured validation error. Built-in evals cannot be edited.

Delete Eval

Archive one of your workspace's authored evals so it no longer lists or runs. Use when retiring a custom grader you no longer need; the slug frees up for re-use and a soft archive keeps the audit trail. Pass the slug. Built-in evals cannot be archived. Returns whether a matching eval was found and archived.

Validate Eval

Dry-run validate a custom eval definition without storing it. Use when building an eval in a form and you want author-time feedback before saving: it runs the same structural checks as creating one (input-schema shape, subject-to-input-field integrity, per-grader config) and returns valid plus a flat list of path + message errors. Never writes anything.

List Eval Runs

List this workspace eval runs, newest first, filterable by eval_slug and status (queued|running|complete|failed|canceled). Use when rendering an eval run history or checking whether a report card has finished. Returns lean rows; fetch eval_run://<id> for the full verdict.

Get Eval Run

Fetch one eval run by id: its status and, once complete, the verdict — the overall score, the pass/partial/fail bucket, and per-case + per-grader scores with rationale + evidence. Use when polling a run started via evals.run or rendering a report card. Returns null for a missing id.

List Grader Kinds