Pick a pattern.
Before the LLM
Bounded observation
The model sees a window of the file and a summarised search, not the whole thing.
Evidence
SWE-agent, table 3: a 100-line window scores 18.0% against 12.7% for the full file; summarised search 18.0% against 12.0% iterative. SWE-bench points the same way: raising the BM25 context from 13k to 50k tokens lifts recall from 29.58 to 51.06 and drops resolution from 1.96% to 1.22% (Claude 2).
SWE-agent, arXiv 2405.15793 ↗ · SWE-bench, arXiv 2310.06770 ↗
Related: Filtered payload between roles · Check on edit, not at the end
Before the LLM
Role separation
One call reasons about the change, a second one writes the edit.
Evidence
Aider, code-editing benchmark, same model in both roles: Claude 3.5 Sonnet 77.4% → 80.5%, GPT-4o 71.4% → 75.2%. On the polyglot leaderboard (June 2025) the split does not always win: o3 (high) alone 81.3%, o3 (high) + GPT-4.1 as editor 78.2%, at lower cost.
Aider architect/editor ↗ · Aider polyglot leaderboard
Related: Filtered payload between roles · Dispatch separated from decision
Before the LLM
Filtered payload between roles
A sub-agent hands back its result, not its transcript.
Evidence
LangChain's default: “the sub-agent's final message is returned as the tool result.” Anthropic describes its research sub-agents as “intelligent filters” that condense what reaches the lead agent. No measurement isolates the effect.
LangChain docs ↗ · Anthropic, multi-agent research system
Related: Role separation · Reviewer without shared context · Bounded observation
Before the LLM
Bounded spend
A spending ceiling is checked before each call, per role.
Evidence
Anthropic: multi-agent systems use about 15× the tokens of a chat, and on BrowseComp token usage alone explains 80% of the performance variance. Spend is the main lever, so it gets a ceiling.
Anthropic, multi-agent research system ↗
Related: Dispatch separated from decision · Progress ledger
Before the LLM
Dispatch separated from decision
Who works next is decided by rules (owner, priority, dependencies), not by a model call.
Evidence
No published benchmark compares deterministic dispatch of scoped tasks against LLM routing. The nearest work (arXiv 2606.28925) routes natural-language prompts, a different problem. The argument is cost: a routing call per step buys nothing when the scope is already known.
No published measurement
Related: Role separation · Bounded spend · Progress ledger · One writer, many readers
After the LLM
Deterministic gate before execution
The model proposes a tool call; a read-only check of the call and the current state can veto it before the write executes.
Evidence
Four pre-execution gates on τ²-bench airline (gpt-4o-mini, pass^1): 29.6% → 42.0%, +12.4 points (74 → 105 of 250, P = 0.0012); a 15-seed replication gives +12.3. It targets what the paper calls silent policy violations: the tool accepts a well-formed call and neither the tool nor the agent's report shows anything wrong.
Reason Less, Verify More ↗ · KDD 2026 workshop (KDD-ETAAI)
Related: Check on edit, not at the end · One writer, many readers
After the LLM
Check on edit, not at the end
Each edit is validated as it is written, not once the agent believes it is done.
Evidence
SWE-agent, table 3: editor with linting 18.0% against 15.0% without. With the check on, 51.7% of trajectories have at least one failed edit, and after the first failure the chance of eventually succeeding drops from 90.5% to 57.2% (§5.2, full SWE-bench, GPT-4 Turbo).
SWE-agent, arXiv 2405.15793 ↗
Related: Deterministic gate before execution · Reproduction test before the fix · Bounded observation
After the LLM
Reproduction test before the fix
A test that reproduces the bug decides which fix is kept.
Evidence
Agentless (SWE-bench Lite, GPT-4o), selecting among candidate patches: majority vote 25.67% → with regression tests 27.00% → with a reproduction test 32.00%. At Google, giving the reproduction test to the repair agent as input: 74% against 57% plausible fixes (17 against 13 of 23 bugs).
Agentless, arXiv 2407.01489 ↗ · Google, arXiv 2502.01821 ↗
Related: Check on edit, not at the end · Reviewer without shared context
After the LLM
Reviewer without shared context
The output is reviewed in a fresh session that never saw how it was produced.
Evidence
Cross-Context Review (30 artifacts, 150 injected errors): a fresh-session review reaches F1 28.6% against 21.7% for a second review in the same session (Holm-adjusted p = 0.004). It is not significantly ahead of a context-aware sub-agent review (23.8%) or a single same-session review (27.1%). One author.
Cross-Context Review, arXiv 2603.12123 ↗
Related: Filtered payload between roles · Reproduction test before the fix
Across turns
One writer, many readers
Only one agent changes state; the others contribute analysis to it.
Evidence
Cognition's position (April 2026): multi-agent systems work best “when writes stay single-threaded and the additional agents contribute intelligence rather than actions.” A vendor position, not a measurement.
Cognition, Multi-Agents: What's Actually Working ↗
Related: Dispatch separated from decision · Deterministic gate before execution
Across turns
First-class human interruption
The run pauses for a person and resumes where it stopped.
Evidence
LangGraph's interrupt() and resume: on resume the interrupted step re-runs from its start, so side effects before the pause should be idempotent. No measurement of the pattern itself.
LangGraph docs, interrupts ↗
Related: Progress ledger · Deterministic gate before execution
Across turns
Progress ledger
A record of whether the team is advancing, with a stall counter that triggers a replan.
Evidence
Magentic-One (§4.1): each iteration the orchestrator answers five questions (done? looping? progressing? who next? what instruction?) and keeps a stall counter; above 2 it replans. The open question is what feeds the counter: the model's judgement, as in the paper, or deterministic signals. Not measured in isolation.
Magentic-One, arXiv 2411.04468 ↗
Related: First-class human interruption · Dispatch separated from decision · Bounded spend