superagentic

Spawn ten agents on one job and they all start on page one — then each invents its own idea of what "done" means.

Two problems, and they have the same cause: a freshly spawned agent has no context. It did not read your orchestration code, it cannot see the other nine, and it will not remember any of this next session. So it needs two things it can only get by asking:

• which unit is mine — nobody else is on it, and if I die it comes back;
• what am I supposed to do with it — the task, what finished looks like, what to hand back.

superagentic is where both live. The orchestrator defines the work and enqueues the units; every worker claims one and is handed the assignment with it. A library, a CLI and an MCP server, in one SQLite file, with no dependencies at all.

uv tool install superagentic                      # the CLI, on your PATH
uvx superagentic demo                             # or run it without installing
uv pip install superagentic                       # or as a library
brew install narimannemo/tap/superagentic         # or via Homebrew

The shape of it

import superagentic as sa
conn = sa.connect("work.db")

# One execution of a fleet. Everything below belongs to it, so afterwards you
# can ask what THIS run did rather than what the database contains.
run = sa.start_run(conn, label="Tomus II extraction")

# The orchestrator, once. This is the part a prompt cannot do:
# the ninth worker, spawned an hour from now, reads the same thing.
sa.define(conn, "extract",
    instructions="Read $path. Record every claim it makes, quoting verbatim.",
    done_when="every claim on the page is recorded, or you have established "
              "there are none",
    returns='{"claims": <int>, "notes": "<string>"}',
    tools="the `xrad` MCP server: record_claim, check_quote")

sa.add(conn, "extract", pages, run=run, meta={"path": "scans/$name.png"})

A kind can also say what a worker must have — separate from what it must do, because a skill it never loaded is not something it can discover halfway through a unit:

sa.define(conn, "extract", instructions=...,
    skills=["xrad-extraction"],
    mcp={"xrad": "xrad serve --db graph.db"},
    context=Path("ontology/glossary.md").read_text())

Then superagentic prompt extract -n 4 generates the spawn prompt from the kind, already telling each worker which skills to load and to fail rather than improvise if it cannot.

Then spawn ten agents with one instruction — claim work and do it — and each of them is handed this:

UNIT: p0189   (kind: extract, id: extract:p0189)

WHAT TO DO
Read scans/p0189.png. Record every claim it makes, quoting verbatim.

USE
the `xrad` MCP server: record_claim, check_quote

DONE WHEN
every claim on the page is recorded, or you have established there are none

HAND BACK
{"claims": <int>, "notes": "<string>"}

Call finish (unit_id=extract:p0189) when done, or fail with a reason.
Do not start any other unit.

When they are finished, sa.results(conn, "extract", run=run) is what they produced, and sa.runs(conn) is every fleet you have ever run:

run                    label                            units  done  failed  workers  elapsed  parallel
20260804-165708-410a * Tomus II extraction                 40     0       0        0      30s      —
20260804-165705-13f3   Tomus I extraction, new prompt      70    70       0        6       4s     5.1x
20260804-165650-24a8   Tomus I extraction                  73    70       3        4       9s     0.8x

parallel is worker-seconds divided by wall-clock — how much concurrency you actually got. 0.8x on a four-worker run means three of them were idle most of the time and the units were too few or too uneven to fill them.

In sixty seconds

uvx superagentic demo

-- 2. three workers claim, and never collide ---------------
   worker-a: page-1, page-2
   worker-b: page-3, page-4
   worker-c: page-5, page-6
   6 units handed out, 6 distinct -- nobody got the same page

-- 3. two finish. the third crashes, holding its work ------
   worker-c: [process dies without reporting anything]

-- 4. its lease expires, and the work comes back -----------
   another worker asks immediately:
     nothing -- still leased
   ...one second later, after the lease expired:
     worker-d picked up page-5  (attempt 2)
   No daemon ran. reclaim() happens on the way into claim().

-- 5. and the dead worker cannot close what it lost --------
   worker-c calls finish on page-5: False
   worker-d calls finish on page-5: True

A lease, not a lock

This is the only hard part of the problem, and every other decision follows from it.

A lock held by a crashed worker is worse than no lock at all. The unit is neither being worked nor available, and nothing in the system can tell a busy worker from a dead one. A lease makes that distinction the passage of time: renew it and you keep the unit, stop renewing and it returns to the pool.

There is no daemon and no cron. reclaim() runs at the top of every claim(), so the next worker asking for work does the cleanup on its way in.

At-least-once, and nothing can do better

Said here rather than in a footnote, because the alternative is you finding out in production:

A worker that is slow rather than dead will have its lease expire, another worker will take the unit, and both will finish it.

No timeout distinguishes those two cases. Two defences, and you want both: heartbeat while you work, so only genuinely stalled units are reclaimed; and make the write at the end idempotent, so a unit done twice converges.

When a lease is lost, finish returns False rather than raising. Handle it — this worker no longer owns the unit and should claim a different one.

The worker loop

while units := sa.claim(conn, "translate", lease=1800):
    for u in units:
        try:
            out = do_the_work(u.name, u.instructions)
            sa.finish(conn, u.unit_id, result=out)
        except Exception as e:
            sa.fail(conn, u.unit_id, note=str(e))

Stages compose without this becoming a scheduler — a finishing worker hands the next stage its units:

sa.finish(conn, u.unit_id, result={"claims": 12},
          then={"audit": [f"claim-{i}" for i in ids]})

Nothing is enqueued if the close failed, so a worker that lost its lease cannot inject work off the back of a unit it no longer owns.

From the shell

claim exits 1 with no output when the queue is dry, so a loop ends by itself. Eight workers, no coordinator:

superagentic add extract --from-file pages.txt

for i in $(seq 1 8); do
  ( while unit=$(superagentic claim extract --json --lease 1800); do
      id=$(echo "$unit" | jq -r '.[0].unit_id')
      name=$(echo "$unit" | jq -r '.[0].name')
      if my-extractor "$name"; then
        superagentic done "$id"
      else
        superagentic fail "$id" --note "extractor exited $?"
      fi
    done ) &
done
wait
superagentic status --who

From an agent

Nine MCP tools, split by who uses them.

The orchestrator — the agent that spawns the fleet — uses define_kind, add_jobs and job_results. It can set up an entire fleet without touching a shell.

Each worker uses claim_job, finish_job, release_job, fail_job, heartbeat_job and job_status.

The tool descriptions carry the protocol, because that is all a worker reads: claim before starting, do what the unit's brief says rather than what you assume the task is, and stop when the queue is empty rather than invent work — which is the failure mode worth designing against, since an agent with nothing to do will reliably find something, and what it finds is usually a unit somebody else has.

{"mcpServers": {"work": {"command": "superagentic",
                         "args": ["serve", "--db", "work.db"]}}}

See MCP.

Watching it run

superagentic dashboard --db work.db          # http://127.0.0.1:8787
superagentic dashboard --out fleet.html      # a static snapshot

Every run, newest first, with what it did — click one to scope every panel to it. 14 left is the same number whether four workers are moving through it steadily or three have died and one is stuck on a poison unit. The dashboard is the difference: throughput over time, what every worker is holding right now and for how long, duration p50 against p95, and a stripe on any unit held past three times the p95 — because "is anyone stuck?" is the question, and a raw duration column does not answer it.

A sidebar switches between projects (each is a database) and runs; selecting one scopes every panel. Served from http.server, CSS and JS inline, SVG drawn by hand. No framework, no build step, nothing fetched.

Read-only, so pointing it at a live run cannot disturb the run. There is an optional access token with sign in and sign out — and because there is no TLS, the server refuses to bind off-loopback unless one is set.

Documentation

What it is not

Not a scheduler. No dependencies between units, no backoff, no cron, one integer of priority. If you need those, run a real queue and keep this for the hand-out.

Not a broker. One SQLite file on one filesystem. Many processes, one box. SQLite over NFS is not safe and this does not pretend otherwise.

Not exactly-once. See above. Nothing is.

It does not do your work or check it. It hands out units and carries your instructions verbatim. Whether the agent followed them is between you and the agent.

Apache-2.0. Contributions welcome — CONTRIBUTING.md says what will and will not be accepted before you spend an evening.