agentfield-use

AgentField subharnesses — offload work instead of doing it inline

A machine with AgentField has one or more control planes — a local one (default http://localhost:8080) and possibly a cloud deployment configured in AgentField Desktop — plus agent nodes installed under ~/.agentfield. Each node exposes reasoners: typed functions you call through the control plane, which routes the call, records the run, and returns the result. You never talk to an agent's own port.

Those installed agents are subharnesses: workers you offload to. They run on smaller, cheaper models, they run in parallel, they run off your context, and every run is recorded on the control plane and watchable in its web UI. A review that would cost you 40k tokens of reading costs you one dispatch and one result.

Vocabulary rule. Internally — commands, API fields, discovery output — everything stays agent / reasoner / node, exactly matching the CLI and API. In what you SAY TO THE USER, call them subharnesses: "your AgentField subharnesses", "the pr-af subharness". Never let the wording cost clarity: if the user needs a command or a node name, give the literal one.

Offload by default

Whenever you have a discrete task — one the user delegated, or one that arose inside your own work — the first question is not "how do I do this?" but "does an installed subharness cover this?" When one does, dispatching to it is the default path, not an option to offer.

Coverage is the test, not size. A small job a subharness covers still goes to the subharness; a large job nothing covers stays with you. This skill carries no list of offloadable task types and cannot: users install new subharnesses at any time, so runtime discovery (§1) is the only source of truth about what your fleet does.

The check is cheap — that is the whole design. One af ls -e (or one capabilities call) answers "does anything cover this?". Discover once per session, remember the roster, and match later tasks against what you learned; re-discover only when something suggests the fleet changed — an install, a start, or a miss you doubt. Consulting this skill on a task nothing covers costs one cheap lookup. That is the trade, deliberately.

The contract, in four parts:

1. Default-offload. Do not ask "would you like me to use an agent?" — check coverage (§1), then dispatch. Work stays inline only when nothing covers it (§1, "No coverage") or the user said to do it yourself.

2. Announce it, with a link. Tell the user what you offloaded, to which subharness, and the run's live URL — built from the resolved server (§0) plus the run_id, URL-encoded: <server>/ui/runs/<run_id>.

   Offloading the PR #42 review to your pr-af subharness — follow along at http://localhost:8080/ui/runs/run_abc123

   The link is for the user to watch in parallel. It never replaces your own monitoring (§4) or your final report (§5). On a cloud target the link only opens if the user's browser can reach that host — print it anyway, the URL shape is identical.

3. The user can always override. "Do it yourself" ends the offload; do the work inline and say so. Overrides are per-request, not permanent.

4. Never silent-wash the offload. If the offloaded run fails, stalls, or comes back empty, report that and ask. Do NOT quietly redo the work inline and present the output as if the subharness produced it. The same rule covers a node that cannot start (§1): never substitute your own work for an agent's without saying so — the user believes their agent ran.

0. Resolve the server first (local vs cloud)

The local and cloud fleets are disjoint: different agents, different versions, different filesystems, different run history. Nothing ever falls back from one to the other on its own. A local server in local mode has no auth; a cloud deployment (and any server with an API key configured) requires X-API-Key: <key> on every request.

Resolution order — stop at the first match:

1. Explicit wins. The user named a server, or AGENTFIELD_SERVER is set in the environment → use that.

2. Read the desktop cloud config. Check every path that applies to this machine — a file that exists but declares no enabled cloud does NOT end the search:

   • macOS: ~/Library/Application Support/agentfield-desktop/settings.json
   • Windows: %APPDATA%/agentfield-desktop/settings.json
   • Linux: ~/.config/agentfield-desktop/settings.json
   • WSL (detect: grep -qi microsoft /proc/version): the Linux path above first, then the Windows side, where the desktop app usually lives: /mnt/c/Users/*/AppData/Roaming/agentfield-desktop/settings.json. A Linux-side file with no cloud key shadowing a Windows file that holds the real cloud config is the common split-brain — the enabled cloud wins, whichever side declares it.

   The first file declaring cloud.enabled: true with a non-empty cloud.serverUrl makes the cloud the target: strip any trailing slash from the URL and take cloud.apiKey as the key. Health-check it (GET <url>/health with X-API-Key).

   • Healthy → use the cloud for everything below.
   • Unreachable → stop and tell the user their cloud control plane is configured but not responding. Do NOT silently fall back to local: work dispatched there lands on a different fleet with different filesystems, which is worse than no dispatch.

3. Otherwise use local: http://localhost:8080.

Then pass the target explicitly on every call: af --server <url> -k <key> (every af command accepts -s/--server and -k/--api-key), or the URL plus -H 'X-API-Key: <key>' for curl. Do not export AGENTFIELD_SERVER yourself to switch targets — a global default leaks into other processes and outlives the task; explicit per-call flags are the contract. If you'd rather not pass -k each time, af auth login --server <url> stores a key per server in ~/.agentfield/credentials.json.

Health-check before the first dispatch: curl -s <server>/health → 200 with {"status":"healthy", ...}. Connection refused on the local target means no control plane is running — the user can open AgentField Desktop, or you can start one in the background (af server blocks, so background it and poll /health). If the resolved target is the configured cloud and this fails, stop and report it — do not retarget local.

The golden path

Five steps, CLI-first. af is installed wherever AgentField is; the HTTP equivalents are further down for what the CLI can't do and as a fallback.

af ls -e -s <server>                                  # 1. what can I offload to?
af call pr-af.review --schema -s <server>             # 2. the exact contract
RUN_ID=$(af call pr-af.review --in '{"pr":42}' --async -s <server>)   # 3. dispatch
af wait "$RUN_ID" --timeout 300 -o json -s <server>   # 4. monitor
# 5. report: result + duration + cost picture + <server>/ui/runs/$RUN_ID

1. Find the subharness — discover, don't guess

Run this once per session and keep the roster; re-run it only when the fleet may have changed (an install, a start, or a miss you doubt).

af ls -e                       # entry-point reasoners only — the callable surface
af ls [query]                  # all reasoners across RUNNING agents (not the install registry)
af agent search "review a pull request"    # BM25-ranked; --agent <id>, --limit N (max 50)
af list                        # installed agents + status (source of truth for INSTALLED)

af agent search hits carry reasoner_id, agent_id, invocation_target, tags, score, and agent_health — everything needed to dispatch with no second lookup. Prefer it over dumping the whole capability payload into context once a box has more than ~20 reasoners.

The durable HTTP discovery endpoint, when you need the whole fleet at once:

curl -s "http://localhost:8080/api/v1/discovery/capabilities?include_input_schema=true"

Reasoner names are .reasoners[].id (NOT .name). Don't assume jq exists (fresh Windows boxes lack it) — parse with what's installed, e.g.:

curl -s "http://localhost:8080/api/v1/discovery/capabilities?include_input_schema=true" -o caps.json
python -c "
import json
for c in json.load(open('caps.json'))['capabilities'] or []:  # null when no agents registered
    print(c['agent_id'], c.get('health_status'), [r['id'] for r in c.get('reasoners',[])])"

Three gotchas:

• The response's invocation_target field uses a colon (agent:reasoner). The execute target uses a dot. Build it yourself: <agent_id>.<reasoner_id>.
• Discovery lists every registered agent, including dead ones — check health_status and only dispatch to "active" agents. Dispatching to an inactive/unknown agent queues work that never runs.
• Installed-but-never-started agents may not appear at all. Discovery lists what is RUNNING; af list is the source of truth for what is INSTALLED. This is the normal first-run state, not an edge case — the desktop app ships swe-planner and pr-af pre-provisioned but deliberately NOT started, because they need API keys the user hasn't entered yet.

Start it before you dispatch — the start attempt is the diagnostic

If the node you need is in af list but absent from discovery, or its health_status isn't active, run af run <name> BEFORE dispatching (it detaches; the agent keeps running after the CLI exits). Do this first, not as a fallback after a failed call: a node blocked on an unset key never registers, so every call to it comes back as the useless agent 'X' not found, while af run names the exact variable and exits 1.

af run reads the encrypted store (~/.agentfield/secrets/*.enc) — the same store that gates startup — so it is the only authoritative check that a node's keys are set. Do not use af doctor or af config <pkg> --list to decide whether a key is configured: doctor reads only the process environment (os.Getenv) and config --list reads the package .env file, so both report a correctly-stored key as ✗ unset, and neither renders require_one_of groups. af secrets ls shows what IS stored but never cross-references manifests, so it can't tell you a required key is missing.

A missing key is a blocking handoff, not a problem to route around. When af run fails with

node swe-planner: missing required environment variables: OPENROUTER_API_KEY (af secrets set OPENROUTER_API_KEY --node swe-planner)

— or, for an alternatives group, at least one of ANTHROPIC_API_KEY or OPENROUTER_API_KEY is required — set one with: … — the value exists only in the user's head. Stop and tell them: the exact variable(s) the error names, the exact af secrets set … --node <name> command copied verbatim from it, and that the same key can be entered in AgentField Desktop → Agents → → Keys. Then wait.

Do NOT retry af run, do NOT dispatch to the node anyway, do NOT substitute a different agent, and do NOT quietly do the job yourself instead — a silent substitution is the worst outcome, because the user believes their agent ran. Never ask the user to paste the secret value into the conversation; the CLI prompt and the desktop form take it directly.

No coverage: offer to build it

Only decide that there is no coverage after completing the health check, capability discovery (including each candidate's description and input schema), and a ranked search for the requested job. Coverage requires a healthy active installed agent whose reasoner description and input schema support that job; a similar name or tag alone is not coverage.

If discovery finds a stopped-but-capable installed agent, explain that it can be started with af run <name>; do not offer a replacement build. If those checks establish that no installed reasoner supports the requested job, say explicitly: "No capable installed agent was found for this job." Then do the work inline yourself (that is the honest fallback — say you are doing it), and offer to build the missing capability: with the agentfield-personal skill when the user wants an agent installed on this machine, or with the agentfield skill for a standalone project repository.

A completed no-coverage result is evidence for the offer, not authorization to create anything. List, inspect, and diagnose-only requests never authorize building an agent. Hand off to a builder skill only when the original request already authorized creating an agent, or when the user explicitly accepts this offer.

2. Fetch the contract

Fetch the exact contract before you dispatch — never guess inputs

Search and discovery tell you a reasoner exists; they do not license a call. Before the first call to any reasoner, read its contract:

af call <node>.<reasoner> --schema        # prints the input schema and exits
af agent agent-summary --id <agent_id>    # all of an agent's reasoners: descriptions + input/output schemas + health + 24h metrics
# single reasoner via MCP: get_reasoner_schema
# or the fleet at once: curl -s "<server>/api/v1/discovery/capabilities?include_input_schema=true"

Read BOTH the description and the input schema, and follow them literally:

• A schema of {"type":"object"} with no properties is NOT "anything goes" — it means the agent registered no schema and the description text is the entire contract. Field names, required-ness, and types stated there are binding (e.g. swe-pro's code_task: goal and an absolute dir are required; model pools are comma-separated strings, not arrays).
• Result semantics live in the description too. Some agents report a failed job in the RESULT (status: "fail") while the execution itself reads succeeded — check the result's own status field, not just the execution's.

Entry points only — undescribed reasoners are internal

Agents register their internal pipeline stages alongside their public flows, and discovery lists all of them. Dispatch ONLY to reasoners that carry the entrypoint tag or a description. A reasoner with no description (e.g. swe-planner's run_* stages) or tagged internal is plumbing invoked by an orchestrator — calling it directly fails or corrupts a run. af ls -e lists tagged entry points; when in doubt, pick the described reasoner whose description names your use case.

3. Dispatch

RUN_ID=$(af call swe-planner.plan --in '{"task":"add rate limiting to the API"}' --async)
# -> bare run_id on stdout; with -o json: {"run_id":"…","status":"accepted"}

What af call does for you: it fetches the schema and validates your input client-side before dispatch, so a bad payload fails locally instead of burning a run. --in also takes @file.json / @file.yaml, and piping JSON to stdin works. --field .path.to.field extracts a single field from a result.

• With af call --in, pass the kwargs at the top level — the CLI wraps them under "input" for you. Over raw HTTP you nest them yourself (§HTTP).
• Always pass --async from a harness. Without it, af call on a TTY auto-tails the run; but a harness's stdout is not a TTY, and there it falls back to the synchronous endpoint with its hard 90s timeout. Async + monitor is the offload path; sync is for quick lookups only.
• If an interactive af call is interrupted it prints Detached. Resume with: af tail <run_id> — the run is still going.

Concurrency — use it

Async dispatch is cheap: fire all independent calls up front, then monitor them together. Do NOT serialize multi-agent work — the whole point of the control plane is managing many subharnesses at once. When a batch of independent jobs arrives (ten PRs to review, five repos to scan), the default is to dispatch the whole batch now and poll as a group — not one-at-a-time. What to know:

• Concurrent calls to the same reasoner are safe when the agent is (e.g. pr-af isolates concurrent reviews per PR). If an agent's docs don't say it's parallel-safe, assume same-target calls may contend on shared state and stagger them; different agents never contend. Some agents serialize ALL executions process-wide (swe-pro queues concurrent code_task calls behind one lock) — the reasoner description says so when known; dispatching more than one heavy call to such a node just builds a queue.
• Each call fans out inside the agent (one review ≈ dozens of sub-executions, several LLM CLI processes). 3–4 heavy runs per node is a sensible ceiling unless the agent documents otherwise.
• Save every run_id you dispatch — you need them to monitor, to report, and for the audit trail. Group related calls with an X-Session-ID header so they're queryable as one batch later.

Check the load before piling on. Every af agent / agentic response carries meta.load: {running_agents, total_agents, active_executions, cpu_cores, recommended_max_concurrent} (the recommendation is CPU-based). Read it before launching more heavy runs — if active_executions >= recommended_max_concurrent, finish or await in-flight work first rather than starting more, and tell the user you're throttling to avoid overloading the machine.

Canary after reconfiguration, then fan out. The one exception to fire-everything-up-front: you just changed a node's runtime config (provider, model, bin path — af secrets set + restart). A misconfigured harness can fail silently — the run reports succeeded with empty results in seconds, and an agent that posts externally (GitHub reviews, Slack, tickets) will publish that garbage under the user's identity, once per dispatched call. So after any config change: send ONE representative call, confirm it did real work (plausible output, a real duration_ms, and nonzero cost in the usage/stats window — not just succeeded), then fan out the rest at full width. This is a gate on the first call after a config change, not a reason to serialize steady-state work.

4. Monitor — pick the retrieval mode

af wait polls /api/v1/agentic/run/:run_id every 2s; default --timeout is 600s. Exit code 2 means TIMEOUT, not failure — the run is still going. Wait again with a longer timeout, or switch to group polling. Exit 1 is a genuinely failed run.

Webhooks are not for you. The execute request body (sync and async) accepts "webhook": {"url": "…", "secret": "…", "headers": {…}}; the response carries webhook_registered (plus webhook_error on async) and the execution status carries webhook_events. That is for services and CI that run an HTTP listener. A coding harness has no listener and must never register one and wait — use wait / tail / poll.

What's in flight right now — no IDs needed: af ps (--agent <name>, --session <id>), or GET /api/v1/executions/active (filters: ?agent_id=, ?session_id=), which returns per-run active_executions, total_executions, started_at, latest_activity.

Several at once: POST /api/v1/executions/batch-status with {"execution_ids": [...]}. Terminal entries embed the FULL result payload — responses can be large (100KB+), so write to a file and parse from there; never pass the response through a command-line argument (Windows caps argv ~32KB).

There is no GET /api/v1/executions list endpoint — use /executions/active for in-flight work and POST /api/v1/agentic/query (body: {"resource":"runs","filters":{"status":"..."},"limit":20}) for history.

Wedge protocol — "running" is not proof of progress

An execution can report running indefinitely after its agent silently dies or deadlocks. Treat a run as suspect when /executions/active shows latest_activity more than ~10 minutes old while active_executions > 0 AND af logs <agent> shows nothing new for that run. (A quiet log alone is not proof — one long LLM completion can be minutes of legitimate silence.) Then:

1. Cancel the WHOLE run, not just the root: POST /api/v1/workflows/<run_id>/cancel-tree (bottom-up, cancels children too). Plain /executions/<id>/cancel cancels ONLY that execution — children keep "running" and must be cancelled individually.
2. Restart the agent if it's wedged: af stop <name> && af run <name>.
3. Re-submit the work — and tell the user it wedged and was re-submitted. A wedged run is a reportable event, not something to paper over.

If the result carries a workspace_handle, you can read the files

Some agents (SWE-AF) mirror the workspace they are building in, so you can open the actual files instead of reasoning from the summary — including uncommitted edits and untracked files that no git push would carry. You do not ask whether this is available and there is nothing to configure: the handle is in the result when it works and absent when it doesn't.

"workspace_handle": {"v":1, "remote":"ssh://host:port"|"dir:/path",
                     "namespace":"...", "key":"<64 hex>", "token":"..."}

furrow is rarely on PATH. AgentField installs it to $AGENTFIELD_HOME/bin/ (default ~/.agentfield/bin/), and a node that ships its own copy keeps it inside the installed package. Resolve it from those; do not try to install it yourself. POSIX sh only — no brace expansion, so the package dirs are spelled out.

os=$(uname -s | tr A-Z a-z)
af_home=${AGENTFIELD_HOME:-$HOME/.agentfield}
furrow_bin() {  # $1 = furrow | furrow-dial
  command -v "$1" 2>/dev/null && return
  for c in "$af_home/bin/$1" \
           "$af_home"/packages/*/bin/"$1"-"$os"-* \
           "$af_home"/packages/*/go/bin/"$1"-"$os"-*; do
    [ -x "$c" ] && { echo "$c"; return; }
  done
}
FURROW=$(furrow_bin furrow) DIAL=$(furrow_bin furrow-dial)

A dir: handle needs only $FURROW; an ssh:// handle needs $DIAL too. If either is missing, say so plainly and carry on from the result — the mirror is fine, this machine just has no client for it.

# ssh:// handle — furrow-dial carries the protocol; nothing else changes
export FURROW_SSH_COMMAND="$DIAL" FURROW_DIAL_TOKEN=<token> FURROW_DIAL_INSECURE=1
FURROW_RECOVERY_KEY=<key> "$FURROW" clone <remote>/<namespace> ./run-workspace --no-watch

# dir: handle (same machine) — clone rejects directory remotes, so pair instead.
# The path is the handle's remote with the "dir:" prefix removed; don't append
# anything to it.
git init -q run-workspace && "$FURROW" --repo run-workspace watch --no-daemon
"$FURROW" --repo run-workspace pair <path> --name <namespace> --key <key>
"$FURROW" --repo run-workspace sync --pull --bootstrap

"$FURROW" --repo run-workspace sync --follow keeps it current while the run works. Read and diff freely. Treat it as a mirror, not a shared drive: it is one-writer, and edits go back as a merge (furrow merge <fork> --check "<cmd>"), so change files between issues or on a fork rather than while the agent writes.

get_workspace_handle re-fetches a handle mid-run: POST /api/v1/execute/<agent>.get_workspace_handle with {"input":{"run_id":"..."}}. {"available": false} means no mirror — carry on without it. Not every build ships this reasoner: check the agent's reasoner list (discovery or agent-summary) before calling it; if it's absent, the node predates the mirror feature and results simply never carry a handle.

5. Report back

Close every offload with: the result, the run's duration_ms, the live URL <server>/ui/runs/<run_id> (run_id URL-encoded), and — when the user would care about spend — the cost picture. Keep the run_id in the transcript; it is the handle for the audit trail and for any follow-up.

If the run failed, timed out, wedged, or returned an empty result on nontrivial input: say so, show what you know (af logs <name>, the error message), and ask how to proceed. Do not fill the gap with your own inline work presented as the subharness's.

Cost: window aggregate, not per-run

Per-execution usage (tokens, provider, model, harness, cost_usd) IS ingested and stored keyed by run, but the only exposed API is the aggregate:

curl -s "http://localhost:8080/api/ui/v1/usage/stats?window=1h"
# window=1h|24h|7d|30d|all (default 24h) -> {totals, by_model, by_provider, by_agent, by_harness}

There is no per-run cost endpoint today. So after finishing a batch of offloaded work, when the user would care, report the cost picture from usage/stats — e.g. the 1h window's by_agent entry for the node you used — stating plainly that it is a window aggregate for that agent, not an exact per-run figure. Never invent a per-run number by dividing or estimating. duration_ms IS exact and per-execution (it is in the execute and status responses). Duration is per-run truth; cost is window truth.

Sessions and multi-call work

• X-Session-ID: <your-id> on execute requests groups multi-turn work; the control plane forwards it to the agent and scopes session memory by it.
• Reuse X-Run-ID across several execute calls to group them into one workflow; each response also returns its run_id.

Agents share state through control-plane memory if you need to pass artifacts around: POST /api/v1/memory/set with {"key": ..., "data": <any>, "scope": "global"} and POST /api/v1/memory/get with {"key": ...} (non-global scopes resolve from the X-Workflow-ID / X-Session-ID / X-Actor-ID headers).

Audit trail

Every execution is recorded — that is part of what makes offloading better than inline work. When provenance matters (or the user asks "what did the agents actually do"), fetch the verifiable-credential chain for a workflow: GET /api/v1/did/workflow/<run_id>/vc-chain (available when DID/VC is enabled), and verify offline with af verify audit.json.

When things fail

MCP (zero-setup)

The control plane serves a built-in MCP server at <server>/mcp (default http://localhost:8080/mcp) — same port, no extra process, on by default. If your harness speaks MCP, this is the fastest way in.

claude mcp add --transport http agentfield http://localhost:8080/mcp
# cloud target (§0):
claude mcp add --transport http agentfield https://<cloud-host>/mcp --header "X-API-Key: <key>"

Other MCP clients: point them at the same streamable-HTTP URL (http://<server>/mcp, transport http). It's stateless JSON-RPC — no session setup. If the server has an API key, pass it as an X-API-Key: <key> header in the client's MCP config.

Five tools are exposed: discover_agents, get_reasoner_schema, execute_reasoner (starts an async run, returns a run_id), get_run, and wait_run. Disable with AGENTFIELD_MCP_ENABLED=false (the route then 404s).

The MCP tools cover the common discover → execute → poll loop. The af CLI and the raw HTTP API remain the full-power path (sessions, streaming, cancel-tree, secrets, load-aware pacing); reach for them when a task needs more than the five tools give you.

HTTP API — where the CLI can't reach

Use these for webhook registration, batch status, memory, cancel-tree, and X-Session-ID/X-Run-ID headers — or as the whole path when af isn't installed. Over raw HTTP, input kwargs are ALWAYS nested under "input" — never raw at the top level. Empty input is {"input": {}}.

# async — the default for real work; returns 202 immediately
curl -s -X POST http://localhost:8080/api/v1/execute/async/swe-planner.plan \
  -H 'Content-Type: application/json' \
  -H 'X-Session-ID: my-batch-1' \
  -d '{"input": {"task": "add rate limiting to the API"}}'
# -> {"execution_id":"...", "run_id":"...", "status":"queued", ...}

# sync — quick lookups only (hard 90s timeout; response carries result + duration_ms)
curl -s -X POST http://localhost:8080/api/v1/execute/swe-planner.plan \
  -H 'Content-Type: application/json' -d '{"input": {"task": "..."}}'

# one execution — poll until status is terminal (succeeded/failed/cancelled/timeout)
curl -s http://localhost:8080/api/v1/executions/<execution_id>
# live progress as Server-Sent Events
curl -s http://localhost:8080/api/v1/executions/<execution_id>/events

# service/CI only: register a webhook at dispatch time
curl -s -X POST http://localhost:8080/api/v1/execute/async/pr-af.review \
  -H 'Content-Type: application/json' \
  -d '{"input":{"pr":42},"webhook":{"url":"https://ci.example/hook","secret":"s3cr3t"}}'

Batched API reads in one round trip: POST /api/v1/agentic/batch with {"operations":[{"id":"op1","method":"GET","path":"/api/v1/agentic/status"}]} (CLI: af agent batch -f operations.json).

Local ops cheat sheet (af CLI)

All commands accept -s/--server <url> and -k/--api-key <key> — required on every invocation when the resolved target is the cloud (§0).

af list                    # installed agents + status
af ls [query]              # search reasoners across running agents (NOT the install registry)
af ls -e                   # only entry-point reasoners — the callable surface
af agent search "<job>"    # ranked reasoner search
af agent agent-summary --id <name>   # full contract: reasoners, schemas, health, 24h metrics
af call <node>.<reasoner> --schema   # input schema only
af call <node>.<reasoner> --in '<json>' --async   # dispatch; prints run_id
af wait <run_id> [--timeout N]       # block until terminal (exit 2 = timeout)
af tail <run_id>           # attach to the live event stream
af ps                      # in-flight runs across all agents (af ps --agent <name>)
af run <name>              # start (detached); af stop <name>
af logs <name>             # agent logs (-f follows; no per-run filter — grep by run_id)
af secrets set KEY [--node <name>]   # store an API key (encrypted; prompts for value)
af secrets ls              # what's configured (values never shown)
af install <git-url>       # install a new agent node

Hard rules

• Offload by default. Any task an installed subharness covers goes to that subharness — whatever its size — announced with its <server>/ui/runs/<run_id> link, not offered as an option and not done inline by habit. Coverage is the test; check it (cheaply, once per session) before doing the work yourself.
• Never silent-wash an offload. A failed, stalled, or empty run is reported and asked about. Never redo it inline and present it as the subharness's work, and never substitute your own work for an agent that can't start.
• Say "subharness" to the user; keep agent / reasoner / node in commands, fields, and anything the user has to type.
• Resolve the server per §0 and pass it explicitly (--server / full URL) on every call. A desktop-configured cloud beats the local default; an unreachable configured cloud is a stop-and-report, never a silent fallback.
• Fetch the reasoner's contract before the first call. A vacuous schema means the description is the contract — follow it literally.
• Dispatch only to entrypoint-tagged or described reasoners. Undescribed or internal-tagged reasoners are pipeline stages — never call them directly.
• Every call goes through the control plane — never POST to an agent's own port. The one exception is a workspace_handle: its ssh:// endpoint is a furrow transport, not the agent's HTTP port, and the per-run token in the handle is what authorizes it. Reading files there is not an agent call.
• Over HTTP, kwargs live under "input" ({"input": {}} when empty). With af call --in, pass them at the top level — the CLI nests them.
• --async + monitor for anything that might exceed a few seconds; sync is for quick lookups only. Independent calls go out together, not one at a time.
• Never register a webhook and wait for it — a coding harness has no listener.
• Only dispatch to agents whose discovery health_status is "active". If it isn't there, af run <name> first — and if that reports a missing required environment variable, stop and hand off to the user with the exact key name and af secrets set command. Never retry it, work around it, or substitute.
• Report duration from duration_ms (exact) and cost from usage/stats (a window aggregate). Never state a per-run cost — there is no such endpoint.
• Don't guess endpoints. The surface above is the contract; if something is missing, ask GET /api/v1/agentic/discover?q=<keyword> before inventing a route.
• Building or modifying an agent (new reasoners, scaffolds, deploys) is the agentfield skill's job — switch to it for that.