Alterspective Agent

Agent …

DEV

We create Flow

Hand the friction over.
Keep the judgement.

Alterspective Agent works inside Microsoft 365, your case management system, and your firm's knowledge. It runs the research, the drafting, the analysis, and the operational round-trips, then brings the work back for you to review and sign.

AI behind the human, earning its moment to help.

For Alterspective staff and connected organisations. Every tool call, file write, and outbound action is recorded and attributed.

What the agent is for

Work where the work already is

Read matters, draft documents, review time and task logs, and update work items inside your case management system and Microsoft 365 — without moving to another screen to feel confident.

Already real Connected tool sources appear as soon as your organisation grants them, per person.
Governed, and able to explain itself

Consequential steps stop and wait for a person. You see the forecast before the act and a receipt after it, and nothing reaches a client system unattributed.

Already real Approval cards sit inline in the conversation, beside the work they belong to.
Shaped to the task

Purpose-built agents for contract review, matter intake, drafting, and operations — each scoped to the tools, knowledge, and approvals that task actually needs.

Already real Operators create, version, and publish agents from Agent Studio.

Why we build it this way

Flow is the state where work absorbs you — the goal is clear, the feedback is immediate, the challenge fits, and effort feels like its own reward. We cannot promise that state. What we can build are the conditions for it.

The Alterspective Flow programme

01A clear goal, stated in your words rather than the system's
02Immediate feedback — every act shows its state change and its receipt
03Context at the point of need, not one screen away
04Interruption spent only where it is earned

Alterspective Agent

Your AI assistant for professional tasks — connect to Microsoft 365, your case management system, and data sources, then ask anything.

Admin

A skill is extra instructions for an agent. It reaches only the agents it is added to, from their next run. Add a skill to an agent in Agent Studio, on the agent's Skills tab.

Detected conversation failures — tool errors, empty results, rejections — recorded by the passive failure log, plus user-reported ones. First soft occurrences (severity info) are hidden unless included. Error text is shown verbatim from the upstream system.

Last seen	Severity	Code	Tool	Count	Status	Detail	Actions

Quality trends

Daily average rubric scores (0–5) from conversation scoring, last 30 days.

Day	Dimension	Avg score	Samples

Calibration

Runs where user feedback and the observer average disagree (👍 against a low score, 👎/flag against a high one) — the rubric-tuning pile, last 30 days.

When	User rating	Observer avg	Run

Stored in the database (DB-over-env). Hot-applicable settings take effect on the next request; restart-required settings (e.g. knowledge source) apply after a service restart.

Approval gate Max loop turns (1–50) System prompt (prepended to every call — leave blank for default) Tool-RAG (narrow tool list by embedding similarity each turn) Tool-RAG top-K Microsoft (Entra) browser sign-in (default off — Keystone SSO is primary; when off the Microsoft button is hidden and Entra API tokens are rejected)

AI gateway (primary LLM). Model + routing tiers apply on the next request; base URL / endpoint are restart-required. The API key stays a secret (not editable here). Failover/routing is owned by the Synapse v2 gateway — there is no consumer-side fallback to configure.

Model (served model id — qwen3-coder:latest → RTX 5090 Ollama, qwen3-coder-next → DGX Spark) Gateway base URL (restart-required) Gateway endpoint (restart-required) Fast model (simple turns — e.g. provider/model-fast; blank = use default) Powerful model (complex turns — e.g. provider/model-powerful; blank = use default) Privacy tier (optional x-privacy-tier header, e.g. local-only; blank = not sent) Quality tier (optional x-quality-tier header, e.g. high; blank = not sent) Task type (optional x-task-type header, e.g. code; blank = not sent) Long-term memory (cross-session recall — restart-required; pgvector needs a database)

Guardrails & Safety. Protect sensitive client information, prevent infinite tool repetition loops, and control live tool routing budgets.

Content filter (PII / DLP redaction before LLM re-injection & emission) Content filter active rules (comma-separated: email, jwt, api_key, au_phone, credit_card; blank = all) Repeat call guard (detects and warns if model repeats the exact same tool call) Turn tool schema budget (max tool schemas offered to LLM per turn; default 12)

Context & Memory Management. Compress long multi-turn conversations into durable summaries to protect LLM context windows.

Context compression Context compression token budget (token threshold that triggers compaction; default 60000)

Sub-Agents & Personal Agent Guardrails. Controls how deep sub-agents can delegate and sets ceiling bounds for user-cloned personal agents.

Max delegation depth (sub-agent nesting level ceiling; default 3) Max delegation fan-out (concurrent sub-agents per round; default 5) Personal agent allowed tools (comma-separated glob patterns; default *) Personal agent max turns (loop turn ceiling for user personal agents; default 12) Personal agent max fan-out (sub-agent fan-out ceiling for user personal agents; default 3) Personal agent token budget (token ceiling for user personal agents; default 400000)

Traffic & Rate Limiting. Bounds user message throughput to prevent accidental burst traffic.

Message send rate limit (requests / minute; default 60) Session creation rate limit (requests / minute; default 20)

Knowledge grounding (RAG). Changing the source is restart-required; limits/threshold apply on the next request.

Knowledge source (restart-required) RAG base URL RAG collection (optional) RAG top-K limit RAG min similarity (0–1; keep low for rank-fusion scores) RAG timeout (ms) (per-search timeout; default 10000 ms = 10s)

RAG token (written to Key Vault server-side — the value is never stored in the app DB or returned)

Status: …

Skill learning. Capture stays a separate switch. Distill drafts candidate skills from captured lessons and never auto-activates them — every skill still needs an operator to approve it. Hot-applies; the live gate is this row, not the env var.

Distill captured lessons (Tier-2 skill_learning_distill_enabled)

Test retrieval. Runs a bounded, rate-limited query against the configured knowledge source. Returns safe metadata only — no secrets.

Test query