Write the objective.
Approve the plan.

01 — Install

Install

AgentGod ships as a single binary with no runtime dependencies. Pick your platform.

# install
curl -fsSL https://get.agentgod.ai | sh

# verify
agentgod --version

irm https://get.agentgod.ai/ps1 | iex

agentgod --version

Set AGENTGOD_API_KEY in your environment before the first run, or you will be prompted for it once and it will be stored in the system keychain.

02 — First run

First run

An objective is a sentence. AgentGod turns it into a plan and waits for you.

agentgod run "compare the three leading vector databases
              on cost, recall, and operational burden"

What you will see

A plan: the tracks it intends to open, the role and tool set for each, and its estimate. Nothing spawns until you accept.

PLAN  5 tracks · est. 4m · 5 agents

  01  scout.pricing     web, archive
  02  scout.benchmarks  web, papers
  03  analyst.recall    python, tabular
  04  analyst.ops       python, docs
  05  editor.synthesis  draft, cite

[a] approve   [e] edit   [q] abort

Approve and it runs. The fleet is torn down automatically when the artifact is written — you do not clean up after it.

03 — Writing objectives

Writing objectives

A good objective states the outcome and the constraints, and leaves the method alone. AgentGod plans better when it is told what done looks like rather than which steps to take.

Declaring a run in a file

For anything recurring, put the objective under version control.

[objective]
goal      = "weekly competitive digest"
deadline  = "15m"
artifact  = "reports/digest-{date}.md"

[limits]
max_agents = 8
budget     = "250k tokens"

[gates]
approve_plan = true   # no spawn before a human accepts

[teardown]
on_complete = "terminate_all"
on_error    = "terminate_all"

04 — Tools & scopes

Tools & scopes

Every agent is sealed at spawn with exactly the tools its track needs. Scopes cannot be widened mid-run — if a track turns out to need more, the orchestrator re-plans and asks you.

[agents.scout]
tools    = ["web.read", "archive.read"]
network  = "allowlist"
allow    = ["*.gov", "*.edu", "docs.*"]

[agents.analyst]
tools    = ["python.sandbox", "fs.read"]
network  = "none"          # cannot reach out at all

[agents.editor]
tools    = ["draft.write", "cite.verify"]
paths    = ["reports/"]     # and nowhere else

05 — Reading the trace

Reading the trace

Every spawn, tool call, finding, and termination is recorded as it happens. Follow it live, or replay it later.

# follow the current run
agentgod watch --trace

# replay a finished run
agentgod trace 7f3a·c21 --verbose

# which agent produced a given claim
agentgod trace 7f3a·c21 --attribute "recall@10 = 0.94"

The trace is append-only. There is no code path in which a privileged action executes without a corresponding entry.

06 — Halting a fleet

Halting a fleet

One signal ends everything. Halting the orchestrator without halting its agents would be the worst of both worlds, so it is not possible.

# stop the current run and every agent under it
agentgod halt

# stop one run by id
agentgod halt 7f3a·c21

# confirm nothing survived
agentgod ps
→ 0 agents live

A halted run still writes its trace and any partial artifact, so you keep the work done up to that point.

07 — Reference

CLI

08 — Changelog

Changelog

AgentGod is early. Interfaces on this page may change before 1.0.