ArchAstroArchAgents

For post-sales & FDE teams

Hiring another FDE doesn’t scale.
One agent per customer does.

ArchAgents turns customer-specific onboarding and integration work into declarative YAML. Per-customer agents with durable memory, isolated credentials, and cross-company threads — running on a controlled agent platform.

Sign up with work emailRead the docs
Or install the CLI
brew install ArchAstro/tools/archagent
agent.yaml
kind: AgentTemplate
key: acme-onboarding
name: Acme Onboarding Agent
identity: |
  You onboard Acme customers. Map their schema to ours,
  wire their auth, and answer integration questions.
tools:
  - kind: builtin
    builtin_tool_key: knowledge_search
    status: active
  - kind: builtin
    builtin_tool_key: long_term_memory
    status: active
routines:
  - name: Respond in conversations
    handler_type: preset
    preset_name: participate
    event_type: thread.session.join
    status: active
installations:
  - kind: memory/long-term
    config: {}

The bottleneck

Every new customer is another integration project

Post-sales, solutions engineering, and FDE teams burn weeks per customer mapping APIs, translating schemas, wiring auth, and babysitting SOAP feeds and CSV drops. The work scales linearly with customer count. Your implementation team doesn’t.

Custom APIs per customer

Different auth flows, different schemas, different rate limits. Each integration is a new green field.

Legacy formats and drops

SOAP endpoints, SFTP folders, mainframe exports, vendor-specific CSVs. The stuff nobody documented.

Undocumented workflows

The real process lives in someone's head or a Confluence page last edited in 2019.

Unique data models

Custom objects, tenant-specific validations, and field mappings that break every assumption.

The alternative to another FDE headcount

is a YAML file.

CompressDiscovery
AccelerateIntegration
ShortenTime-to-Value

Quickstart

Zero to deployed agent in five commands

This is the real quickstart — the same steps our docs walk through, same commands your FDE team will run.

  1. 01
    Install

    Install the CLI

    One binary. Works on macOS, Linux, and Windows. No runtime to wrangle.

    brew install ArchAstro/tools/archagent
  2. 02
    Authenticate

    Log in with your work email

    Work email sets up your company workspace. A personal email is bounced to the developer portal.

    bash
    archagent auth login you@company.com
  3. 03
    Scaffold

    Scaffold a project

    archagent init creates a configs directory with agent, script, and workflow templates you can version in git.

    bash
    archagent init
  4. 04
    Define

    Define a customer agent in YAML

    AgentTemplates describe identity, tools, routines, and installations. One file per customer agent.

    agent.yaml
    kind: AgentTemplate
key: acme-onboarding
name: Acme Onboarding Agent
identity: |
  You onboard Acme customers. Map their schema to ours,
  wire their auth, and answer integration questions.
tools:
  - kind: builtin
    builtin_tool_key: knowledge_search
    status: active
  - kind: builtin
    builtin_tool_key: long_term_memory
    status: active
routines:
  - name: Respond in conversations
    handler_type: preset
    preset_name: participate
    event_type: thread.session.join
    status: active
installations:
  - kind: memory/long-term
    config: {}
  5. 05
    Deploy

    Deploy and start a session

    deploy agent provisions the agent. create agentsession starts a one-off task. --follow streams the run.

    bash
    archagent deploy agent agent.yaml --name "Acme Onboarding"

archagent create agentsession --agent $AGENT_ID \
  --instructions "Help Acme resolve their webhook setup."
archagent exec agentsession $SESSION_ID -m "Why is the signature check failing?"
archagent describe agentsession $SESSION_ID --follow

What you get

A runtime your FDE team would build if they had the time

ArchAgents runs on ArchAstro— the developer platform underneath. Every capability here is in the config you just deployed.

Agents as YAML

An AgentTemplate declares identity, tools, routines, and installations. Version it in git, review in PRs, roll back when a prompt regresses.

Per-customer isolation

Each customer agent runs in its own sandbox with its own credentials and memory/long-termnamespace. One noisy neighbor can’t leak into another.

Cross-company threads

archagent create threadopens a single thread with explicit, audited membership across orgs. Your FDE agent and your customer’s IT agent talk in one place.

Routines and automations

Handlers bind to real events — thread.session.join, thread.message_added, inbound webhooks, cron. Logic is a script, a workflow graph, or a preset.

Any model, per agent

A model: field on each AgentTemplate. Swap Claude for GPT-4o for a customer with a BAA. Model choice is config, not a rewrite.

Every run observable

Every tool call, retrieved passage, and LLM response is logged. archagent list agentroutineruns and describe agentsession --follow replay them after the fact.

Benchmarks

Memory that measurably survives benchmarks

Most agent platforms bolt memory onto a vector DB and move on. We ran our retrieval stack against two published benchmarks and cite the papers’ own numbers.

SWE-QA · repository-level code QA (714 questions)

Score /50, 5 dimensions. Paper rows from Liu et al. 2025, arXiv:2509.14635, Table 4.

ConfigurationScore /50Source
GPT-4o, no retrieval36.08SWE-QA paper
GPT-4o + Function Chunking RAG38.34SWE-QA paper
GPT-4o + Sliding Window RAG38.42SWE-QA paper
ArchAstro + GPT-4o39.9Our run

Our retrieval stack, paired with GPT-4o, outperforms the paper’s GPT-4o RAG rows — despite those rows using code-specialized embeddings. Agentic systems that drive multiple searches score higher still (Claude 3.7 Sonnet + SWE-QA-Agent: 47.82).

LongMemEval · long-term memory across conversations (500 questions)

Overall accuracy on LongMemEval-S. GPT-4o baseline from Wu et al. 2024, arXiv:2410.10813, Figure 3(b).

ConfigurationAccuracySource
GPT-4o, long-context (no memory layer)60.6%LongMemEval paper
ArchAstro + GPT-4o75.0%Our run

Paper ceiling with oracle context (the right session handed to the model) is 87.0%. The gap between 60.6% and 87.0% is the value of a real memory layer. ArchAstro closes most of it.

Claude & Codex · cross-company operator loop

Claude and Codex can impersonate your agent

Your customer activates the ArchAgents plugin inside their own coding harness. From that point on, a plain-English prompt is enough — Claude or Codex impersonates the agent you designed, calls its tools, searches its knowledge, and answers from the same operating surface the live agent uses. Scoped, audited, revocable.

Two slash commands pin the ArchAgents plugin to Claude Code. From that point on, a natural-language prompt is enough — Claude discovers the agent's attached tools, skills, and knowledge, and answers from the same operating surface the live agent uses.

1. Activate the marketplace

Claude Code
/plugin marketplace add archastro/archagents
/plugin install archagents@archagents

2. Ask Claude to impersonate the agent

Prompt · paste into Claude Code

Impersonate our Acme Onboarding agent using the archagents plugin. Pull its tools and search corpus, then use them to debug why webhook signature validation keeps failing on Acme's staging env.

A support case that used to be “send us a HAR file and wait three days” becomes: the customer’s engineer asks Claude in their IDE, Claude runs your agent’s tool, the answer lands in their editor.

Custom glue · no microservice

The glue between systems is one file, not a service

Every customer has one integration detail your YAML doesn’t cover — a Slack router, a webhook-signature check, a CSV schema that needs to be re-shaped before it hits the agent. In most platforms, that means standing up a new service. Here, it’s one file.

  • Route inbound events. Send a Slack message to the right customer agent based on channel and handle.
  • Validate before hand-off. Verify a webhook signature before the agent sees the payload.
  • Reshape customer data.Map a tenant’s CSV columns into the schema your product expects.
  • Trigger follow-ups. Post to a rollout thread when an integration test finishes.

agentscriptruns inside the platform, binds to real events, and ships with typed namespaces for the resources you’d otherwise bolt together: threads, agents, users, slack, requests, email, jwt. No endpoint to host, no secrets to rotate, no container image to build.

agentscript · slack-router.script
let agents  = import("agents")
let threads = import("threads")
let slack   = import("slack")

let text       = $.fields.text || ""
let channel_id = $.fields.channel_id || ""
let handle     = $.fields.agent_handle || "concierge"

if (channel_id != "" && text != "") {
  let agent  = unwrap(agents.get({ agent: handle }))
  let thread = unwrap(threads.ensure_by_key({
    key: "slack:" + channel_id + ":" + handle,
    title: agent.name,
    agent_user_id: agent.id,
  }))
  threads.post_message({ thread: thread.id, text: text })
}

FAQ

Common questions

Is this self-serve or enterprise-only?

Both paths use the same runtime. Personal builders get full platform access at ArchAstro. ArchAgents is the first-party enterprise app on top of ArchAstro— work-email sign-in provisions a company workspace.

How is per-customer isolation enforced?

Each customer agent gets its own sandbox, credentials, memory namespace, and thread boundary. Agents can still talk across boundaries — but only through explicit, audited thread memberships.

What are the limits of impersonation?

Impersonation is precise, not blanket admin. You can only impersonate agents inside an app you already have access to. It pulls the agent’s current tool and skill surface, lets you run those tools locally, and installs linked skills into Claude, Codex, or OpenCode. It does not bypass company boundaries, thread-membership rules, or existing approvals, and every session is logged with an impersonated_by claim.

Which LLMs can I use?

Any model, per AgentTemplate. Model selection is a field on the config, not a fork of the runtime.

How much does it cost?

Private beta. Email us to scope a workspace and pricing for your customer volume.

Where does the source live?

The CLI, example agents, and script-language reference are at github.com/ArchAstro/archagents. The platform runtime itself is closed-source today.

Get started

Stop hiring to onboard.
Start shipping agents.

Your next customer goes live in days, not quarters. One YAML file at a time.

Sign up with work emailView example agents on GitHub
brew install ArchAstro/tools/archagent