Role definition
Job-to-be-done, inputs, outputs, escalation rules, refusal cases. Versioned in your repo — a role is a piece of code, not a saved prompt.
Capability · Agix Agents
Autonomous teammates that get sharper for your team over time.
Agix Agents are autonomous teammates with defined roles, governed tool access, and memory that compounds. They take work off your team — researching, drafting, reviewing, escalating — and learn from every chat. The same agent runtime that powers our internal C-suite agents (Director, Atlas, Maven, Architect, Ops, Hunter, Librarian) is what your team gets, tuned to your roles.
What an Agix Agent is
An autonomous teammate with a role, a memory, a critic, and an eval set. Unlike a copilot — which lives inside another tool and helps a human act — an agent works on its own surface, manages its own work, and gets demonstrably sharper for your team each time it runs.
- Custom roles tailored to your team's workflows, data, and voice
- Memory that compounds — every chat session makes the agent sharper for your team's preferences and rules
- A critic pass on every output before it ships to a human reviewer
- Tool access governed by your policies; high-risk actions stay human-approved
- Eval harness on every prompt and policy change — agents that demonstrably improve, with proof
- Provider-portable across Anthropic, OpenAI, and enterprise endpoints
The stack we install
Six layers — built on the same runtime as Agix's internal agents — that turn a role description into an operating teammate.
Memory store
Per-workspace long-term memory of lessons, exemplars, and team preferences. Schema is yours; surfaced as in-context examples on the next run.
Tool gateway
Governed access to your systems — CRM, queues, drafts, calendars. Read paths broad by default; write paths narrow and human-approved unless explicitly opened.
Critic loop
A second-pass agent reviews every output before it ships to a human. Disagreements between writer and critic feed memory — the agent gets sharper at its own job each cycle.
Eval harness
Canned regression suite over real cases from your workflow. Prompt and policy changes are gated on green; an agent never ships an update that worsens its own score.
Runtime + observability
The same runtime that powers Agix's internal C-suite agents — per-run traces, per-tenant cost protection, refusal logs, audit trail. Operated by your team from day 91.
Roles you can commission
Start from a template that already works for our team and tune it to yours. Or specify a custom role and we'll build it on the same runtime.
Researcher
Crawls your second-brain and the public web, returns sourced briefs with citations, files findings into your knowledge system on its own.
Drafting Analyst
Takes structured inputs (CRM record, support ticket, claim) and produces a first-draft document — proposal, response, summary — tuned to your tone and rules.
Ops Coordinator
Watches a queue, triages incoming work, routes or auto-handles based on explicit escalation rules. Files an audit trail with every decision.
Account Assistant
Lives next to a single account — knows the history, monitors signals, drafts the next outreach, hands off to the human when stakes go up.
Agent Packs — commissioned multi-agent rosters
One agent is a teammate. A pack is a whole function. We commission packs around a specific workflow — a roster of scoped agents with explicit handoffs and a closed feedback loop. The first pack we'll commission is below; more land as we build them with clients.
B2B SaaS Social-Media Pack
Seven agents running an end-to-end social-media operation — trend-listening, ideation, drafting, brand-critique, scheduling, engagement triage, and weekly insights that feed the next cycle's listening prompts.
Built around the buyer persona of a 25-person growth-stage B2B SaaS marketing team running LinkedIn primary, X reactive, and YouTube depth. See the worked example below for the per-agent breakdown and the closed-loop diagram.
See the worked example →Looking for off-the-shelf agents rather than a commissioned pack? See the productized Agix Agents Pack — Solo Operator, Founder, and All-Access tiers.
The recursive learning loop — what makes Agix Agents different
Most AI agents stop improving the moment the engagement ends. Agix Agents are engineered to get better every time your team uses them. Four mechanisms make that real.
Chat-feedback ingestion
Every thumb, edit, and comment becomes a labeled trace. The agent's prompt and exemplar set evolve toward what your team actually wants — without retraining the model.
Long-term agent memory
Per-workspace memory of lessons, exemplars, and team preferences. Surfaced as in-context examples on the next run, so the agent doesn't relearn your conventions on every chat.
Reflexion critic pass
A second agent reviews every output before it ships. Surfaced disagreements become memory entries — the agent gets better at its own job each cycle.
Closed-loop eval
Prompt and policy changes are gated by a regression suite. An agent never ships an update that worsens score on canned cases. Improvement is the only direction.
What your team owns at handoff
The agent is yours from day 91 — code, memory, evals, policies, dashboards. Agix steps back; the runtime keeps running.
The role config
Versioned in your repo. Your team modifies the prompt, refusal rules, and escalation thresholds without us in the loop.
The memory store
Yours. Schema documented, accessible directly, exportable any time. The agent's accumulated knowledge does not live in a vendor cloud.
The eval set
Canned regressions seeded from real cases in your workflow. Your team adds new cases as the agent encounters new territory; deploys are gated on green.
Tool policies
Per-action permissioning — who can read what, who can write what, what stays human-only. Reviewable, versionable, auditable.
Deployment + dashboards
Infra-as-code, observability stack, per-tenant cost dashboard, runbook library. Your ops team runs it; we step back.
When this tier fits
Four buyer situations where commissioning an Agix Agent is the right move.
"We have a recurring research/draft/triage task that doesn't fit into an existing tool."
Agents are the right shape when the work needs its own surface — not embedded inside Salesforce, Word, or your admin UI. A new role, with its own runtime.
"We want something that gets better over time per team."
The recursive learning loop is what makes this tier distinct. Memory compounds; the critic-loop sharpens; the eval-set keeps regression at bay. The agent your team is using in month six is materially better than the one we handed off.
"We need autonomy with explicit human-approval gates."
High-stakes work. Default policy is read-broad, write-narrow, human-approve. The audit trail is mandatory, not optional, and every escalation has a recorded reason.
"We need a role we don't have headcount for."
Researcher, drafting analyst, ops coordinator, account assistant — agent roles take work off your team in shapes where adding a person is wrong or impossible.
When to pick something else
Agents are powerful, but not the right tier for every situation. Here's how to tell.
If the work happens inside an existing tool
Start with AI Copilots instead. Same intelligence, embedded where work already happens — Slack, Salesforce, your admin UI — rather than living in its own surface.
Explore AI Copilots →If the AI is for your customers, not your team
Start with AI-Powered Products. Per-tenant safety, brand-aligned voice, rollout flags — the shape that ships to thousands of customers, not a small operations team.
Explore AI-Powered Products →If you don't have a knowledge substrate yet
Build a Second-Brain first. Agents query the same surface — without it they have nothing reliable to reason over, and you'll rebuild the same knowledge layer per agent.
Explore Second-Brain Knowledge Systems →Worked example: a single role — Research Analyst for a boutique investment firm
An illustrative scenario showing how a single-role commission ships. The shape mirrors how Agix commissions agent roles; the client and specifics are composite.
The brief
“Our analysts spend the first two hours of every morning scanning the same news feeds, filings, and earnings transcripts to produce a daily sector brief for the partners. The work is high-skill but repetitive, and the briefs get edited heavily by whoever's on senior duty that week — claims softened, benchmarks added, certain speculative angles cut. We want an autonomous teammate that produces the first draft of each daily brief overnight, learns from the senior editor's changes, and files into our existing research system so analysts step into the day reviewing instead of composing.”
The shape we'd build
A Research Analyst agent that watches the firm's curated news + filings + transcript feeds, the internal research archive, and the firm's house style guide. Every overnight cycle it drafts one sector brief per active coverage area, runs each through a critic pass that checks against editorial policy (no Fed-timing speculation, always benchmark energy against XLE, never make compliance-sensitive forward-looking claims), and files the draft into the firm's existing research-management queue with citations back to every source it pulled from.
Memory accumulates the senior editor's patterns — the kinds of claims they soften, the benchmarks they prefer per sector, the structural moves they make on weak openers — and surfaces those as in-context examples on the next overnight run. Tool access is read on the feeds and archive; write into the draft queue; never publishes externally. Every brief logs its citations so a partner reviewing a call months later can reconstruct exactly which sources grounded each claim.
What the firm owns when we step back
The role config in their repo. The memory store and its schema. The eval set seeded from real editor revisions — extended by the firm's editorial team as the agent encounters new sectors. The tool policies. The dashboards. The agent runs in the firm's environment from day one and keeps running there; Agix steps back, the runtime stays.
Worked example: a commissioned pack — B2B SaaS Social-Media Pack
Seven agents running an end-to-end social-media operation for a 25-person growth-stage B2B SaaS marketing team. LinkedIn primary, X reactive, YouTube depth. A team of scoped agents with explicit handoffs and a closed feedback loop — not one mega-agent.
1. Listener
Scans X, LinkedIn, Reddit, customer Slack, and support tickets for resonant topics in your market.
In: keyword list + RSS / API feeds. Out: daily Trend Brief — 5–10 ranked topics with citations and a one-line ‘why now?’ for each.
Hands off to → Ideator
2. Ideator
Turns each topic into 3–5 post angles in your brand voice, mapped to platform format.
In: Trend Brief + your Voice Profile (memory). Out: Angle Set per topic, ranked by predicted resonance.
Hands off to → Drafter
3. Drafter
Writes the full piece — LinkedIn carousel copy + slide outline, X thread, YouTube interview prompts — from the chosen angle.
In: approved Angle + Voice Profile + your past top-10 performing pieces (memory). Out: Draft Bundle with markdown copy and an asset-requirements line.
Hands off to → Brand Critic
4. Brand Critic
Reviews drafts for voice, compliance, factual accuracy, and reputational risk. A social-media-tuned variant of our internal Curator agent, applied to your brand voice rubric.
In: Draft Bundle + Brand Voice Rubric + recent industry context. Out: Review Report with risk flags, voice score, and inline suggestions. Reflections feed back into the recursive-learning loop.
Hands off to → Human approver → Scheduler
5. Scheduler
Owns the 30-day content calendar across LinkedIn, X, YouTube. Picks optimal publish times. Auto-staggers to prevent feed collisions.
In: approved Draft Bundles + calendar constraints + historical engagement-time data. Out: Calendar State (what publishes when, on which channel, with what creative assets).
Hands off to → Platform publishers → Engagement Triager (once live)
6. Engagement Triager
Watches comments, mentions, and DMs across all channels. Auto-responds to FAQs in your voice. Escalates real conversations to humans with full context.
In: platform webhooks + Response Playbook + escalation rules. Out: auto-responses sent + Escalation Tickets in CRM for human-required threads.
Hands off to → Analyst
7. Analyst
Pulls performance data daily, identifies what's working, writes a weekly Insights Report — and a structured Next-Cycle Brief that feeds the Listener's prompt for the following week.
In: platform analytics + engagement-triager logs + Calendar State. Out: Insights Report (markdown + charts) + Next-Cycle Brief.
Hands off to → Listener — the cycle closes
The closed loop
The Analyst's weekly Next-Cycle Brief feeds directly into the Listener's prompt for the following week — the cycle closes. Two streams of signal feed the recursive-learning loop at the same time: the Brand Critic's reflection events surface drift in voice, compliance, and risk patterns; the Analyst's insights surface what kinds of posts actually resonated. Both flow into long-term memory and prompt evolution per agent, so the pack you're running in month six is materially sharper than the one we commissioned on day one.
What your team owns at handoff
Each of the seven agent role configs, versioned in your repo. The shared memory store with its schema. The eval set, seeded from your real post history and extended by your team as the pack encounters new territory. The tool policies — read on platform APIs, write into your CMS draft queue, never auto-publish without a human approver in the middle. The dashboards. The pack runs in your environment from day one and keeps running there; Agix steps back, the runtime stays.
Talk through an Agix Agent for your team
Start with a Discovery conversation. We'll talk through the role you have in mind, the policies it would operate under, and what shipping it into your stack looks like.
Looking for the productized off-the-shelf catalog — the named agents you can install on your machine or run in our cloud, rather than the engagement-services architecture? Browse the Agents Pack catalog.
The brief
“We produce case studies, thought-leadership posts, product launches, and reactive commentary across three channels. We struggle with maintaining brand voice across multiple authors, keeping a consistent cadence, and turning engagement signal into the next content cycle. We want a roster of agents — not one assistant — where each agent owns a step and hands off cleanly to the next.”
The seven-agent roster