Multi-Agent Case Studies and Eval Farms · AI Engineering

Three 2026 reference systems, the software team capstone, and a fenced, resumable eval farm in Go.

Three reference cases

Factor	Anthropic Research	MetaGPT / ChatDev	OpenClaw / Moltbook
Topology	Supervisor-worker	SOP role decomposition	Population-scale substrate
Key result	+90.2% over single-agent Opus 4	MacNet: >1000 agents via DAGs	~2.3M agent accounts in days
Cost signal	15x tokens per query	Per-role budget	Per-interaction pricing
Pattern to copy	Fresh-context subagents	Structured artifact handoffs	Prompt-injection threat model
Verification	Explicit verifier role	QA Engineer	In-network validators
Framework	Claude Agent SDK or LangGraph	CrewAI or MetaGPT	Custom substrate + MCP + A2A
Pick for	Research and knowledge tasks	Engineering tool chains	Network-effect social products

Sources

Common patterns

Orchestrator + workers: every case has a supervisor or network effect.
Handoff contracts: artifacts between roles follow a schema.
Verification: a first-class role in every case.
Scaling: topology + substrate, not just more agents.
Effort scaling: simple is 1 agent with 3-10 tool calls. Complex is 10+ subagents.
Rainbow deploy: keep old runtime versions alive for in-flight agents.
Prompt injection: treat agent profiles and cross-agent messages as untrusted input.

Sources

phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md

Map a design to a case

def map_to_case(d: Design) -> str:

if d.task_type == "population" or d.user_facing_network:

return "openclaw_moltbook"

if d.task_type == "engineering" or d.roles_distinct:

return "metagpt_chatdev"

return "anthropic_research"research and the default

print_case(case_key)patterns, framework, citation

print_landscape()FRAMEWORK_LANDSCAPE rows

Sources

phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/code/main.py

Software team build steps

Task board file-backed JSONL of typed messages. Agents subscribe to tags.
Architect emits one plan_request with a DAG of subtasks and interfaces.
Coders each claims one subtask in a fresh git worktree add branch.
Merge coordinator three-way merges the N branches into a staging branch.
Reviewer cannot approve diffs it authored. Routes review_feedback to the coder.
Tester runs the suite in a clean sandbox. Failures loop back to the owner.
Accounting one span per role-boundary message with payload size and model.
Post-mortem a handoff-failure histogram for each failed issue.

Sources

phases/19-capstone-projects/10-multi-agent-software-team/docs/en.md

Eval farm architecture

Nodes and connections

Node	Role	Kind
cases.json	id, prompt, expected, response	client
Partition	FNV-1a into shards
farm.json	under ledger.lock	store
run	Parallel pool, bounded waves	process
ClaimShard	lock, load, acquire, persist	process
WorkShard	exact match outside the lock	process
CompleteShard	reload, Submit while locked	process
Report	coverage and receipts	client

From	To	Connection
cases.json	Partition
Partition	farm.json	InitFarm
farm.json	run
run	ClaimShard	worker process
ClaimShard	WorkShard	cases
WorkShard	CompleteShard	result
CompleteShard	Report

Sources

Eval farm commands

Command	What it does
`go build -o /tmp/eval-farm .`	Build the reference binary in `solution`
`init --store DIR --input FILE --shards N`	Bind dataset hash and shard count
`run ... --shards N --workers W`	Launch bounded process waves
`worker --store DIR --worker-id NAME`	Claim, evaluate and submit one shard
`status --store DIR`	Validate and report the snapshot
`demo --input FILE`	Print crash and recovery snapshots
`--crash-after-claim`	Exit 86 after the durable claim
`--all --solution --strict`	Grade the reference with `project_test.py`

Sources

Eval farm contracts

Task	Code
Stage 1, stable shard per ID	`Partition(ids []string, shards int)`
Hash an ID	`h := fnv.New32a()`
Stage 2, fence a lease	`Acquire(l, worker, now, ttl)`
Stage 3, accept one result	`Submit(l, worker, version, now, result)`
Stage 4, bounded pool	`Parallel(ctx, ids, workers, run)`
Lease fields	`Shard, Worker, Version, Until, Result, Done`
Errors	`ErrInvalid`, `ErrConflict`, `ErrLimit`

Sources

Lease and receipt rules

Acquire: a done or live lease is a conflict

if l.Done || (l.Worker != "" && now < l.Until) {

l.Until = now + ttl

l.Version++

Submit: check owner and version first

if l.Worker != worker || l.Version != version {

a done lease accepts only the same result again

if l.Result == result {

a first completion must come before expiry

if now >= l.Until {

Sources

Defaults and limits

Setting	Default	Note
`--shards`	4	Fixed partition count, range 1..1024
`--workers`	2	Concurrent processes, range 1..64
`--lease-ms`	30000	Claim duration
`--now-ms`	-1	Uses the wall clock
`--delay-ms`	0	Range 0..60000
`--store`	farm-data	Same-host directory
Snapshot size	4 MiB	Whole-run snapshot
Result size	1 MiB	Largest accepted result

Sources

Failure modes

Symptom	Cause	Fix
Worker exits 86	Intentional crash after claim	Run again at the expiry boundary
Run exits 1 with complete:false	Remaining leases still owned	Invoke run again after expiry
Old worker gets a fencing error	Replacement committed first	Expected. Snapshot stays unchanged
Run exits 1 on a changed dataset	Dataset hash differs	Use a different store
Subtasks cannot be implemented	Plan too vague	Plan template with explicit interfaces
Reviewer approves a bug	Reviewer false-approve	Injected-bug probe, tune the prompt
Viral prompt spreads to agents	Prompt injection at scale	Treat cross-agent messages as untrusted
Research agents hallucinate	No explicit verifier role	Add a verifier, about 20-30% of tokens

Sources

Numbers to remember

Token usage: explains 80% of BrowseComp variance.
MAST: failure rates of 41-86.7% on real multi-agent systems.
OpenClaw: 247k GitHub stars in ~4 months.
Capstone eval: 50 SWE-bench Pro issues against one Sonnet 4.7 baseline.
Rubric: 25 pass@1, 20 speedup, 20 review quality, 20 tokens, 15 coordination.
Farm sample: four records score 3/4. Two waves launch four processes.

Sources

Questions

Why does multi-agent research win?: Each subagent gets a fresh context window.
Which benchmark?: SWE-bench Pro. Verified is contaminated.
What does the capstone measure?: Whether multi-agent wins per dollar, not only whether it works.
Can a late correct worker overwrite?: No. old/v1 fails when new/v2 completed the shard.
Does recovery at 109 steal the lease?: No. It reports a pending run and exits 1.

Sources

Original source

---
id: multi-agent-case-studies
title: Multi-Agent Case Studies and Eval Farms
code: MF
subtitle: Three 2026 reference systems, the software team capstone, and a fenced, resumable eval farm in Go.
hue: 210
sources: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota, phases/19-capstone-projects/10-multi-agent-software-team, projects/distributed-eval-farm
source_hash: 75cf32b6d4c0f5e0
---

## Three reference cases
type: comparison
from: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md, phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/code/main.py

| Factor | Anthropic Research | MetaGPT / ChatDev | OpenClaw / Moltbook |
|---|---|---|---|
| Topology | Supervisor-worker | SOP role decomposition | Population-scale substrate |
| Key result | +90.2% over single-agent Opus 4 | MacNet: >1000 agents via DAGs | ~2.3M agent accounts in days |
| Cost signal | 15x tokens per query | Per-role budget | Per-interaction pricing |
| Pattern to copy | Fresh-context subagents | Structured artifact handoffs | Prompt-injection threat model |
| Verification | Explicit verifier role | QA Engineer | In-network validators |
| Framework | Claude Agent SDK or LangGraph | CrewAI or MetaGPT | Custom substrate + MCP + A2A |
| Pick for | Research and knowledge tasks | Engineering tool chains | Network-effect social products |

## Common patterns
type: concepts
from: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md

- **Orchestrator + workers**: every case has a supervisor or network effect.
- **Handoff contracts**: artifacts between roles follow a schema.
- **Verification**: a first-class role in every case.
- **Scaling**: topology + substrate, not just more agents.
- **Effort scaling**: simple is 1 agent with 3-10 tool calls. Complex is 10+ subagents.
- **Rainbow deploy**: keep old runtime versions alive for in-flight agents.
- **Prompt injection**: treat agent profiles and cross-agent messages as untrusted input.

## Map a design to a case
type: annotated-code
lang: python
from: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/code/main.py

```python
def map_to_case(d: Design) -> str:
    if d.task_type == "population" or d.user_facing_network:
        return "openclaw_moltbook"
    if d.task_type == "engineering" or d.roles_distinct:
        return "metagpt_chatdev"
    return "anthropic_research"        # research and the default
print_case(case_key)                   # patterns, framework, citation
print_landscape()                      # FRAMEWORK_LANDSCAPE rows
```

## Software team build steps
type: flow
from: phases/19-capstone-projects/10-multi-agent-software-team/docs/en.md

1. **Task board**: file-backed JSONL of typed messages. Agents subscribe to tags.
2. **Architect**: emits one `plan_request` with a DAG of subtasks and interfaces.
3. **Coders**: each claims one subtask in a fresh `git worktree add` branch.
4. **Merge coordinator**: three-way merges the N branches into a staging branch.
5. **Reviewer**: cannot approve diffs it authored. Routes `review_feedback` to the coder.
6. **Tester**: runs the suite in a clean sandbox. Failures loop back to the owner.
7. **Accounting**: one span per role-boundary message with payload size and model.
8. **Post-mortem**: a handoff-failure histogram for each failed issue.

## Eval farm architecture
type: architecture
from: projects/distributed-eval-farm/README.md, projects/distributed-eval-farm/API.md, projects/distributed-eval-farm/stages/02-fence-each-shard-lease/docs/en.md

```arch
node | cases | 1,1 | cases.json | id, prompt, expected, response | client
node | part | 2,1 | Partition | FNV-1a into shards |
node | store | 3,1 | farm.json | under ledger.lock | store
node | run | 4,1 | run | Parallel pool, bounded waves | process
node | claim | 4,2 | ClaimShard | lock, load, acquire, persist | process
node | work | 3,2 | WorkShard | exact match outside the lock | process
node | done | 2,2 | CompleteShard | reload, Submit while locked | process
node | rep | 1,2 | Report | coverage and receipts | client
edge | cases | part |
edge | part | store | InitFarm
edge | store | run |
edge | run | claim | worker process
edge | claim | work | cases
edge | work | done | result
edge | done | rep |
```

## Eval farm commands
type: command-table
from: projects/distributed-eval-farm/README.md, projects/distributed-eval-farm/API.md, projects/distributed-eval-farm/stages/02-fence-each-shard-lease/docs/en.md

| Command | What it does |
|---|---|
| `go build -o /tmp/eval-farm .` | Build the reference binary in `solution` |
| `init --store DIR --input FILE --shards N` | Bind dataset hash and shard count |
| `run ... --shards N --workers W` | Launch bounded process waves |
| `worker --store DIR --worker-id NAME` | Claim, evaluate and submit one shard |
| `status --store DIR` | Validate and report the snapshot |
| `demo --input FILE` | Print crash and recovery snapshots |
| `--crash-after-claim` | Exit 86 after the durable claim |
| `--all --solution --strict` | Grade the reference with `project_test.py` |

## Eval farm contracts
type: task-table
from: projects/distributed-eval-farm/solution/stage1.go, projects/distributed-eval-farm/solution/stage2.go, projects/distributed-eval-farm/solution/stage3.go, projects/distributed-eval-farm/solution/stage4.go, projects/distributed-eval-farm/solution/types.go

| Task | Code |
|---|---|
| Stage 1, stable shard per ID | `Partition(ids []string, shards int)` |
| Hash an ID | `h := fnv.New32a()` |
| Stage 2, fence a lease | `Acquire(l, worker, now, ttl)` |
| Stage 3, accept one result | `Submit(l, worker, version, now, result)` |
| Stage 4, bounded pool | `Parallel(ctx, ids, workers, run)` |
| Lease fields | `Shard, Worker, Version, Until, Result, Done` |
| Errors | `ErrInvalid`, `ErrConflict`, `ErrLimit` |

## Lease and receipt rules
type: annotated-code
lang: go
from: projects/distributed-eval-farm/solution/stage2.go, projects/distributed-eval-farm/solution/stage3.go

```go
// Acquire: a done or live lease is a conflict
if l.Done || (l.Worker != "" && now < l.Until) {
l.Until = now + ttl
l.Version++
// Submit: check owner and version first
if l.Worker != worker || l.Version != version {
// a done lease accepts only the same result again
if l.Result == result {
// a first completion must come before expiry
if now >= l.Until {
```

## Defaults and limits
type: defaults
from: projects/distributed-eval-farm/solution/main.go, projects/distributed-eval-farm/API.md, projects/distributed-eval-farm/README.md, projects/distributed-eval-farm/stages/03-accept-idempotent-completion-receipts/docs/en.md

| Setting | Default | Note |
|---|---|---|
| `--shards` | 4 | Fixed partition count, range 1..1024 |
| `--workers` | 2 | Concurrent processes, range 1..64 |
| `--lease-ms` | 30000 | Claim duration |
| `--now-ms` | -1 | Uses the wall clock |
| `--delay-ms` | 0 | Range 0..60000 |
| `--store` | farm-data | Same-host directory |
| Snapshot size | 4 MiB | Whole-run snapshot |
| Result size | 1 MiB | Largest accepted result |

## Failure modes
type: troubleshooting
from: projects/distributed-eval-farm/README.md, projects/distributed-eval-farm/API.md, projects/distributed-eval-farm/stages/04-run-a-bounded-local-worker-pool/docs/en.md, phases/19-capstone-projects/10-multi-agent-software-team/docs/en.md, phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md

| Symptom | Cause | Fix |
|---|---|---|
| Worker exits 86 | Intentional crash after claim | Run again at the expiry boundary |
| Run exits 1 with complete:false | Remaining leases still owned | Invoke run again after expiry |
| Old worker gets a fencing error | Replacement committed first | Expected. Snapshot stays unchanged |
| Run exits 1 on a changed dataset | Dataset hash differs | Use a different store |
| Subtasks cannot be implemented | Plan too vague | Plan template with explicit interfaces |
| Reviewer approves a bug | Reviewer false-approve | Injected-bug probe, tune the prompt |
| Viral prompt spreads to agents | Prompt injection at scale | Treat cross-agent messages as untrusted |
| Research agents hallucinate | No explicit verifier role | Add a verifier, about 20-30% of tokens |

## Numbers to remember
type: concepts
from: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md, phases/19-capstone-projects/10-multi-agent-software-team/docs/en.md, projects/distributed-eval-farm/README.md

- **Token usage**: explains 80% of BrowseComp variance.
- **MAST**: failure rates of 41-86.7% on real multi-agent systems.
- **OpenClaw**: 247k GitHub stars in ~4 months.
- **Capstone eval**: 50 SWE-bench Pro issues against one Sonnet 4.7 baseline.
- **Rubric**: 25 pass@1, 20 speedup, 20 review quality, 20 tokens, 15 coordination.
- **Farm sample**: four records score 3/4. Two waves launch four processes.

## Questions
type: qa
from: phases/16-multi-agent-and-swarms/25-case-studies-2026-sota/docs/en.md, phases/19-capstone-projects/10-multi-agent-software-team/docs/en.md, projects/distributed-eval-farm/README.md, projects/distributed-eval-farm/stages/03-accept-idempotent-completion-receipts/docs/en.md

- **Why does multi-agent research win?** Each subagent gets a fresh context window.
- **Which benchmark?** SWE-bench Pro. Verified is contaminated.
- **What does the capstone measure?** Whether multi-agent wins per dollar, not only whether it works.
- **Can a late correct worker overwrite?** No. old/v1 fails when new/v2 completed the shard.
- **Does recovery at 109 steal the lease?** No. It reports a pending run and exits 1.